coder/coderd/provisionerjobs.go

300 lines
8.7 KiB
Go
Raw Normal View History

feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
package coderd
import (
"context"
"database/sql"
"encoding/json"
"errors"
"fmt"
"net/http"
"strconv"
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
"time"
"github.com/go-chi/render"
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
"github.com/google/uuid"
"cdr.dev/slog"
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
"github.com/coder/coder/database"
"github.com/coder/coder/httpapi"
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
)
// ProvisionerJobStaus represents the at-time state of a job.
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
type ProvisionerJobStatus string
const (
ProvisionerJobPending ProvisionerJobStatus = "pending"
ProvisionerJobRunning ProvisionerJobStatus = "running"
ProvisionerJobSucceeded ProvisionerJobStatus = "succeeded"
ProvisionerJobCancelled ProvisionerJobStatus = "canceled"
ProvisionerJobFailed ProvisionerJobStatus = "failed"
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
)
type ProvisionerJob struct {
ID uuid.UUID `json:"id"`
CreatedAt time.Time `json:"created_at"`
StartedAt *time.Time `json:"started_at,omitempty"`
CompletedAt *time.Time `json:"completed_at,omitempty"`
Error string `json:"error,omitempty"`
Status ProvisionerJobStatus `json:"status"`
WorkerID *uuid.UUID `json:"worker_id,omitempty"`
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
}
type ProvisionerJobLog struct {
ID uuid.UUID `json:"id"`
CreatedAt time.Time `json:"created_at"`
Source database.LogSource `json:"log_source"`
Level database.LogLevel `json:"log_level"`
Output string `json:"output"`
}
// Returns provisioner logs based on query parameters.
// The intended usage for a client to stream all logs (with JS API):
// const timestamp = new Date().getTime();
// 1. GET /logs?before=<timestamp>
// 2. GET /logs?after=<timestamp>&follow
// The combination of these responses should provide all current logs
// to the consumer, and future logs are streamed in the follow request.
func (api *api) provisionerJobLogs(rw http.ResponseWriter, r *http.Request, job database.ProvisionerJob) {
follow := r.URL.Query().Has("follow")
afterRaw := r.URL.Query().Get("after")
beforeRaw := r.URL.Query().Get("before")
if beforeRaw != "" && follow {
httpapi.Write(rw, http.StatusBadRequest, httpapi.Response{
Message: "before cannot be used with follow",
})
return
}
var after time.Time
// Only fetch logs created after the time provided.
if afterRaw != "" {
afterMS, err := strconv.ParseInt(afterRaw, 10, 64)
if err != nil {
httpapi.Write(rw, http.StatusBadRequest, httpapi.Response{
Message: fmt.Sprintf("unable to parse after %q: %s", afterRaw, err),
})
return
}
after = time.UnixMilli(afterMS)
} else {
if follow {
after = database.Now()
}
}
var before time.Time
// Only fetch logs created before the time provided.
if beforeRaw != "" {
beforeMS, err := strconv.ParseInt(beforeRaw, 10, 64)
if err != nil {
httpapi.Write(rw, http.StatusBadRequest, httpapi.Response{
Message: fmt.Sprintf("unable to parse before %q: %s", beforeRaw, err),
})
return
}
before = time.UnixMilli(beforeMS)
} else {
before = database.Now()
}
if !follow {
logs, err := api.Database.GetProvisionerLogsByIDBetween(r.Context(), database.GetProvisionerLogsByIDBetweenParams{
JobID: job.ID,
CreatedAfter: after,
CreatedBefore: before,
})
if errors.Is(err, sql.ErrNoRows) {
err = nil
}
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprintf("get provisioner logs: %s", err),
})
return
}
if logs == nil {
logs = []database.ProvisionerJobLog{}
}
render.Status(r, http.StatusOK)
render.JSON(rw, r, logs)
return
}
bufferedLogs := make(chan database.ProvisionerJobLog, 128)
closeSubscribe, err := api.Pubsub.Subscribe(provisionerJobLogsChannel(job.ID), func(ctx context.Context, message []byte) {
var logs []database.ProvisionerJobLog
err := json.Unmarshal(message, &logs)
if err != nil {
api.Logger.Warn(r.Context(), fmt.Sprintf("invalid provisioner job log on channel %q: %s", provisionerJobLogsChannel(job.ID), err.Error()))
return
}
for _, log := range logs {
select {
case bufferedLogs <- log:
default:
// If this overflows users could miss logs streaming. This can happen
// if a database request takes a long amount of time, and we get a lot of logs.
api.Logger.Warn(r.Context(), "provisioner job log overflowing channel")
}
}
})
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprintf("subscribe to provisioner job logs: %s", err),
})
return
}
defer closeSubscribe()
provisionerJobLogs, err := api.Database.GetProvisionerLogsByIDBetween(r.Context(), database.GetProvisionerLogsByIDBetweenParams{
JobID: job.ID,
CreatedAfter: after,
CreatedBefore: before,
})
if errors.Is(err, sql.ErrNoRows) {
err = nil
}
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprint("get provisioner job logs: %w", err),
})
return
}
// "follow" uses the ndjson format to stream data.
// See: https://canjs.com/doc/can-ndjson-stream.html
rw.Header().Set("Content-Type", "application/stream+json")
rw.WriteHeader(http.StatusOK)
rw.(http.Flusher).Flush()
// The Go stdlib JSON encoder appends a newline character after message write.
encoder := json.NewEncoder(rw)
for _, provisionerJobLog := range provisionerJobLogs {
err = encoder.Encode(convertProvisionerJobLog(provisionerJobLog))
if err != nil {
return
}
}
ticker := time.NewTicker(250 * time.Millisecond)
defer ticker.Stop()
for {
select {
case <-r.Context().Done():
return
case log := <-bufferedLogs:
err = encoder.Encode(convertProvisionerJobLog(log))
if err != nil {
return
}
rw.(http.Flusher).Flush()
case <-ticker.C:
job, err := api.Database.GetProvisionerJobByID(r.Context(), job.ID)
if err != nil {
api.Logger.Warn(r.Context(), "streaming job logs; checking if completed", slog.Error(err), slog.F("job_id", job.ID.String()))
continue
}
if job.CompletedAt.Valid {
return
}
}
}
}
func (api *api) provisionerJobResources(rw http.ResponseWriter, r *http.Request, job database.ProvisionerJob) {
if !job.CompletedAt.Valid {
httpapi.Write(rw, http.StatusPreconditionFailed, httpapi.Response{
Message: "Job hasn't completed!",
})
return
}
resources, err := api.Database.GetWorkspaceResourcesByJobID(r.Context(), job.ID)
if errors.Is(err, sql.ErrNoRows) {
err = nil
}
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprintf("get provisioner job resources: %s", err),
})
return
}
apiResources := make([]WorkspaceResource, 0)
for _, resource := range resources {
if !resource.AgentID.Valid {
apiResources = append(apiResources, convertWorkspaceResource(resource, nil))
continue
}
agent, err := api.Database.GetWorkspaceAgentByResourceID(r.Context(), resource.ID)
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprintf("get provisioner job agent: %s", err),
})
return
}
apiAgent, err := convertWorkspaceAgent(agent)
if err != nil {
httpapi.Write(rw, http.StatusInternalServerError, httpapi.Response{
Message: fmt.Sprintf("convert provisioner job agent: %s", err),
})
return
}
apiResources = append(apiResources, convertWorkspaceResource(resource, &apiAgent))
}
render.Status(r, http.StatusOK)
render.JSON(rw, r, apiResources)
}
func convertProvisionerJobLog(provisionerJobLog database.ProvisionerJobLog) ProvisionerJobLog {
return ProvisionerJobLog{
ID: provisionerJobLog.ID,
CreatedAt: provisionerJobLog.CreatedAt,
Source: provisionerJobLog.Source,
Level: provisionerJobLog.Level,
Output: provisionerJobLog.Output,
}
}
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
func convertProvisionerJob(provisionerJob database.ProvisionerJob) ProvisionerJob {
job := ProvisionerJob{
ID: provisionerJob.ID,
CreatedAt: provisionerJob.CreatedAt,
Error: provisionerJob.Error.String,
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
}
// Applying values optional to the struct.
if provisionerJob.StartedAt.Valid {
job.StartedAt = &provisionerJob.StartedAt.Time
}
if provisionerJob.CompletedAt.Valid {
job.CompletedAt = &provisionerJob.CompletedAt.Time
}
if provisionerJob.WorkerID.Valid {
job.WorkerID = &provisionerJob.WorkerID.UUID
}
switch {
case provisionerJob.CancelledAt.Valid:
job.Status = ProvisionerJobCancelled
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
case !provisionerJob.StartedAt.Valid:
job.Status = ProvisionerJobPending
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
case provisionerJob.CompletedAt.Valid:
if job.Error == "" {
job.Status = ProvisionerJobSucceeded
} else {
job.Status = ProvisionerJobFailed
}
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
case database.Now().Sub(provisionerJob.UpdatedAt) > 30*time.Second:
job.Status = ProvisionerJobFailed
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
job.Error = "Worker failed to update job in time."
default:
job.Status = ProvisionerJobRunning
feat: Add provisionerdaemon to coderd (#141) * feat: Add history middleware parameters These will be used for streaming logs, checking status, and other operations related to workspace and project history. * refactor: Move all HTTP routes to top-level struct Nesting all structs behind their respective structures is leaky, and promotes naming conflicts between handlers. Our HTTP routes cannot have conflicts, so neither should function naming. * Add provisioner daemon routes * Add periodic updates * Skip pubsub if short * Return jobs with WorkspaceHistory * Add endpoints for extracting singular history * The full end-to-end operation works * fix: Disable compression for websocket dRPC transport (#145) There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing. This is just tracking some experimentation to fix that race condition ## Run results: ## - Run 1: peer test failure - Run 2: peer test failure - Run 3: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45 ``` status code 412: The provided project history is running. Wait for it to complete importing!` ``` - Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176 ``` workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` - Run 5: peer failure - Run 6: Pass ✅ - Run 7: Peer failure ## Open Questions: ## ### Is `dRPC` or `websocket` at fault for the data race? It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it. From `dRPC`'s perspective, the callstack looks like this: - [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253) - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65) - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128) - with finally the data race happening here: ```go // AppendFrame appends a marshaled form of the frame to the provided buffer. func AppendFrame(buf []byte, fr Frame) []byte { ... out := buf out = append(out, control). // <--------- ``` This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame. Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73 However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race. In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly). ### Why does cloning on `Read` fail? Get a bunch of errors like: ``` 2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF 2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0 ``` # UPDATE: We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now: - Run 1: ✅ - Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338 - Run 3: ✅ - Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168 - Run 5: ✅ * fix: Remove race condition with acquiredJobDone channel (#148) Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83 __Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs. __Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up. * fix: Bump up workspace history timeout (#149) This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32 Looking at the timing of the test: ``` t.go:56: 2022-02-02 21:33:21.964 [DEBUG] (terraform-provisioner) <provision.go:139> ran apply t.go:56: 2022-02-02 21:33:21.991 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.050 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.090 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.140 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.195 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running t.go:56: 2022-02-02 21:33:22.240 [DEBUG] (provisionerd) <provisionerd.go:162> skipping acquire; job is already running workspacehistory_test.go:122: Error Trace: workspacehistory_test.go:122 Error: Condition never satisfied Test: TestWorkspaceHistory/CreateHistory ``` It appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout. Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here. In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that. Co-authored-by: Bryan <bryan@coder.com>
2022-02-03 20:34:50 +00:00
}
return job
}
func provisionerJobLogsChannel(jobID uuid.UUID) string {
return fmt.Sprintf("provisioner-log-logs:%s", jobID)
}