Description
This is an issue we ran into CI, it's a rare occurrence but there's a non-zero chance it can happen to our users as well.
Problem: When a new workspace build is created for an existing workspace (e.g. transition into deleted) a new (unique-ish) name is generated for that job via namesgenerator.GetRandomName(1). Normally this should not be an issue, but we have a unique constraint on (workspace_id, name):
ALTER TABLE ONLY workspace_builds
ADD CONSTRAINT workspace_builds_workspace_id_name_key UNIQUE (workspace_id, name);This results in a non-zero chance that both the create job and delete job receive the same name, resulting in the conflict. For very long lived workspaces, the chance of hitting this issue becomes greater. I did a local test that reproduced this after running workspace start/stop 200 times.
The CI run that hit this issue: https://github.com/coder/coder/runs/7399324482?check_suite_focus=true
2022-07-18T22:52:32.7769558Z workspaces_test.go:331:
2022-07-18T22:52:32.7770135Z Error Trace: /home/runner/work/coder/coder/coderd/workspaces_test.go:331
2022-07-18T22:52:32.7770615Z Error: Received unexpected error:
2022-07-18T22:52:32.7772201Z POST http://127.0.0.1:42617/api/v2/workspaces/9a172a87-b686-4a93-a286-70d473562375/builds: unexpected status code 500: Internal error inserting workspace build.
2022-07-18T22:52:32.7773646Z Error: execute transaction: insert workspace build: pq: duplicate key value violates unique constraint "workspace_builds_workspace_id_name_key"
2022-07-18T22:52:32.7774246Z Test: TestWorkspaceByOwnerAndName/Deleted
2022-07-18T22:52:32.7774730Z Messages: delete the workspace
2022-07-18T22:52:32.7775638Z t.go:81: 2022-07-18 22:49:15.883 [DEBUG] (provisionerd) <github.com/coder/coder/provisionerd/provisionerd.go:370> (*Server).closeWithError closing server with error {"error": null}
2022-07-18T22:52:32.7776288Z --- FAIL: TestWorkspaceByOwnerAndName/Deleted (6.92s)
2022-07-18T22:52:32.7776505Z
2022-07-18T22:52:32.7776777Z make: *** [Makefile:175: test-postgres] Error 1
Questions
- Is there a good reason to keep unique names for all workspace builds (and all transitions)?
- Should we try a new name on conflict?