Shard a slow job so it runs in parallel
A long serial job split across parallel shards, with the same test count and a measured wall-clock.
The ticket
Split a slow pipeline job into parallel shards.
Acceptance criteria
- The total test count is identical before and after
- Wall-clock time is recorded either side
- Shards are balanced rather than split arbitrarily
- A failure in any shard fails the job
What lands as proof
Matching test counts plus the wall-clock difference, so speed is not bought by silently skipping work.
Why teams defer it
- Sharding can silently drop tests if the split is wrong, and a faster green build hides it.
- Balancing shards needs per-test timing data that nobody collects by default.
Questions
- What does the agent actually change?
- The ticket is scoped to one outcome: split a slow pipeline job into parallel shards. Work that serves that outcome is in scope, and anything outside it is left for a separate ticket, so the pull request stays reviewable.
- How do I know the work is done?
- The pull request carries the evidence, not only the diff. Here that means wall-clock before and after with an identical test count, so a reviewer can confirm the result without reproducing the work locally.
- How much oversight does this need?
- The run stops once the plan is written. Somebody reads the approach and approves it before any code exists, which is the cheapest moment to redirect the work.