Optimize a pipeline for cost, time, performance, and reliability

Measure, change, verify

Baseline pipeline-duration percentiles, per-task duration, and queued/running-job pressure. Identify the measured constraint, change one lever, then compare duration, failures, queueing, and cost/quota consumption. Repeat rather than stacking unmeasured changes.

The levers

Lever	Effect
Caching	Accelerate cross-run regeneration; cache absence must not make the job fail
Parallelism	Split independent work across jobs when dependencies, agents, and capacity allow
Path filters	Do not build at all for documentation-only commits
Matrix	One definition, many legs, throttled by `maxParallel`
Pipeline artifacts	Carry required outputs between jobs; unlike a cache, their absence is a failure
Agent choice	Balance clean/stateless isolation against deliberate state reuse and ownership

Caching and the key

A Cache@2 primary key can combine stable tool/OS discriminators with a lockfile path; the task hashes the file contents. Caches are immutable. A fixed overbroad key can retain outdated contents, while an overspecific primary key reduces exact-hit reuse. restoreKeys can return a partial hit when the exact key misses, so a commit SHA defeats cross-commit exact reuse but not every possible fallback hit.

yaml

- task: Cache@2
  inputs:
    key: 'npm | "$(Agent.OS)" | package-lock.json'
    restoreKeys: |
      npm | "$(Agent.OS)"
    path: $(Pipeline.Workspace)/.npm

Measure restore/save overhead and test the cache-miss path. Use a pipeline artifact—not a cache—for files a downstream job requires.

Capacity, agents, and cost

Azure DevOps YAML jobs run in parallel by default only when no dependency blocks them and agents plus organization-level parallel-job capacity are available. Declare dependsOn for required ordering and bound matrix fan-out with maxParallel.

Private projects can use a monthly-limited Microsoft-hosted free tier or paid parallel-job capacity with no monthly limit. Self-hosted capacity is also licensed by concurrent jobs, while its machines, patching, and isolation remain your responsibility. Microsoft-hosted pools supply a clean VM. Managed DevOps Pools add a managed choice: stateless supplies a fresh agent for each job; stateful reuse can preserve files/packages but needs an explicit security and cost decision. Benchmark VM size, storage, region, standby, and state instead of assuming bigger or warmer is better.

Primary sources

Optimising a pipeline

Optimize a pipeline for cost, time, performance, and reliability

Measure, change, verify

The levers

Caching and the key

Capacity, agents, and cost