Problem
A PR pipeline run takes 74–103m of wall clock. The structural floor is ~54m:
Initial Artifact Build 15m (serial prefix -- every test job waits on it)
slowest test job 39m
----
floor 54m
Everything above that floor is queue time (one run fans out ~26 jobs / ~587 job-minutes; one observed run accumulated 271 job-minutes of queue, with individual jobs waiting 20m).
The floor itself is set by the slowest single test job, and two jobs were tied at the top:
|
shards |
test time |
max shard |
ideal (balanced) |
| Integration |
7 |
130.8m |
38.9m (MainSuite 1a) |
26.5m |
| Postman |
11 |
147.8m |
37.8m (GraphQL) |
22.8m |
Because they are nearly equal, rebalancing either one alone moves wall clock by ~1 minute — the other immediately becomes the tail. They have to be fixed together.
Two further findings from measuring:
GraphQLTests is a single 25.6m collection. With ~9.4m of fixed cost per Postman shard, no amount of regrouping gets Postman below ~35m. It is the binding constraint on the entire pipeline.- The
build-classes artifact is uploaded on every build and consumed by nothing. It costs ~0.5m on the serial critical path. The only restore-classes: true caller is deploy-javadoc, which no workflow invokes.
Scope
- Rebalance integration shards on measured per-class time rather than class count (class count is a poor proxy — one shard held 298 classes/31m, another 52 classes/14m).
- Rebalance Postman groups on measured newman time, and shard
GraphQLTests by folder rather than splitting the 518KB collection file.
- Merge the Postman micro-groups that are ~85% fixed overhead.
- Delete the dead
build-classes artifact.
- Hold total job count flat so this does not add queue pressure.
Out of scope
Queue time — the dominant term above the 54m floor. Reducing it needs either runner capacity or fewer jobs, and is worth measuring separately once the tail is down.
Acceptance
- Slowest test job drops from ~39m to ~26m.
- Total job count unchanged.
- No test is dropped, duplicated, or silently skipped.
Problem
A PR pipeline run takes 74–103m of wall clock. The structural floor is ~54m:
Everything above that floor is queue time (one run fans out ~26 jobs / ~587 job-minutes; one observed run accumulated 271 job-minutes of queue, with individual jobs waiting 20m).
The floor itself is set by the slowest single test job, and two jobs were tied at the top:
Because they are nearly equal, rebalancing either one alone moves wall clock by ~1 minute — the other immediately becomes the tail. They have to be fixed together.
Two further findings from measuring:
GraphQLTestsis a single 25.6m collection. With ~9.4m of fixed cost per Postman shard, no amount of regrouping gets Postman below ~35m. It is the binding constraint on the entire pipeline.build-classesartifact is uploaded on every build and consumed by nothing. It costs ~0.5m on the serial critical path. The onlyrestore-classes: truecaller isdeploy-javadoc, which no workflow invokes.Scope
GraphQLTestsby folder rather than splitting the 518KB collection file.build-classesartifact.Out of scope
Queue time — the dominant term above the 54m floor. Reducing it needs either runner capacity or fewer jobs, and is worth measuring separately once the tail is down.
Acceptance