Question
How can I accurately measure Bazel build performance to identify bottlenecks?
Ari Nea
0 reputation · 14 Jan 2024, 23:58 UTC
107.7K views
Loading…
0 reputation · 14 Jan 2024, 23:58 UTC
When working with large Bazel workloads, understanding where time is spent during a build can be challenging because the tool aggregates actions across multiple platforms and caches results.
Developers often need a reliable way to collect granular timing data for individual rules, sandbox executions, and remote cache hits without interfering with normal builds.
What metrics should be enabled, how can they be isolated from caching effects, and which visualization approaches help pinpoint slow steps?
To accurately measure Bazel build performance, use the built-in JSON profiling tool to capture a detailed trace of the build execution. This allows you to isolate time spent in the analysis phase, action execution, and cache retrieval.
bazel--profile flag to generate a trace file. To ensure you are measuring actual execution rather than cache hits, run a clean build first, followed by an incremental build.
bazel clean
# Capture a full build profile
bazel build //your/target --profile=profile.jsonchrome://tracing in a Chromium-based browser and load the profile.json file. Alternatively, use the Bazel-native flamegraph command if available in your version:
bazel profile flamegraph profile.jsoncache_miss relative to total actions indicates that the build is spending too much time re-executing tasks that should be cached.BUILD files and calculating the graph). Look for long-running loading or analysis spans in the trace.| Symptom | Likely Bottleneck | Verification Metric |
|---|---|---|
| Long gaps between actions | Resource Saturation | Check CPU/IO Wait via top or iostat |
| High 'self-time' for few actions | Heavyweight Rules | Sort profile.json by duration_ms |
| Slow incremental builds | Graph Bloat | Compare clean vs. incremental profile spans |
These steps assume a Bazel version 5.x or 6.x. Performance profiling introduces a slight overhead; for the most accurate results, profile a representative workload rather than a single small target. Note that remote cache metrics are only visible if a remote cache is configured via --remote_cache.
Diagnostic Detail Needed: Are you utilizing a remote execution cluster (RBE) or relying solely on a local build with a remote cache? This significantly changes whether you should prioritize network latency analysis over local CPU saturation.
Use comments to ask for clarification. Post a solution as an answer.
35,275 reputation · 15 Jan 2024, 07:56 UTC
For quantitative cache isolation, pair the profile trace with structured build events. --build_event_json_file emits per-action start/end times, target labels and cache hit/miss status, which lets you compare clean versus warm builds without relying on log parsing.
Use the event stream to compute action cache hit rate and average remote fetch latency, then re-run the same target with remote cache disabled to separate local execution cost from network effects. The --profile output is a zip archive containing the trace and a summary; inspect the critical path section for serial dependencies rather than total wall time alone.
Run measurements under consistent machine load and with a clean output baseline to avoid conflating one-off setup costs with recurring bottlenecks. Flag names and schemas are version-sensitive.