Repository navigation
CoderWorkspacePoolMember: recognize a template-wide last_used_at write in a one-member pool #349
Description
Problem
A CoderWorkspacePoolMember treats a later last_used_at as use, except for a template-wide write. It recognizes a template-wide write in two ways: the value is within 5 s of the template's updated_at, or another unclaimed member of the pool already moved its baseline to the same value.
When a template schedule update with update_workspace_last_used_at runs, and the template changes again before the next 60 s poll, and no other member of the pool saw the first write (for example a one-member pool), the member marks itself Used. The pool releases it, and one warm workspace leaks.
This is the fail-safe direction: the controller never deletes a workspace that someone may use. But it costs one workspace per affected member.
Possible fix
Also accept as a witness any other workspace of the template in Coder with the same last_used_at (one list call, only when the template time does not match).
Trigger
Before the PR 12 activation docs of #153, or a user report.
Found by Codex on #343 (comment 4268119093).
Refs #153
Found in review of #343 (thread 4268119093); the plan resolves the ambiguity toward Used, which can leak one warm workspace but never deletes possible work.
Refs #153
Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: medium
Activity
ThomasK33 commentedon Oct 14, 2026
Two more cases from the final assessment of #343, where real use can be counted as "not used" (the side that leads toward replacement, not toward keep). Both windows are short: continued use bumps
last_used_atagain at the next usage flush, about 60 s later, so only a single short use can be missed.updated_atmoves on every template edit (for example a version promote), not only on a schedule update withupdate_workspace_last_used_at. A use flush within ±5 s of an unrelated template edit is treated as a template-wide write and not counted as use.nowto every workspace in a flush batch (coderd/workspacestats/tracker.go,flush). The member compares whole seconds, so a sibling whose baseline came from the same flush or the same second makes this member's real use look like a template-wide write.Also for the docs at activation:
last_used_atis best-effort. Port-forward or raw tunnel connections with no session do not bump it under the usage experiment, and a failed flush drops updates.Gate: decide and document before PR 12 registers the reconciler.
Generated with
xum• Model:anthropic:claude-opus-5-5• Thinking:medium