Debugging Remote Cache Hits for Remote Execution

Report an issue View source Nightly · 7.4 . 7.3 · 7.2 · 7.1 · 7.0 · 6.5

This page describes how to check your cache hit rate and how to investigate cache misses in the context of remote execution.

This page assumes that you have a build and/or test that successfully utilizes remote execution, and you want to ensure that you are effectively utilizing remote cache.

Checking your cache hit rate

In the standard output of your Bazel run, look at the INFO line that lists processes, which roughly correspond to Bazel actions. That line details where the action was run. Look for the remote label, which indicates an action executed remotely, linux-sandbox for actions executed in a local sandbox, and other values for other execution strategies. An action whose result came from a remote cache is displayed as remote cache hit.

For example:

INFO: 11 processes: 6 remote cache hit, 3 internal, 2 remote.

In this example there were 6 remote cache hits, and 2 actions did not have cache hits and were executed remotely. The 3 internal part can be ignored. It is typically tiny internal actions, such as creating symbolic links. Local cache hits are not included in this summary. If you are getting 0 processes (or a number lower than expected), run bazel clean followed by your build/test command.

Troubleshooting cache hits

If you are not getting the cache hit rate you are expecting, do the following:

Ensure re-running the same build/test command produces cache hits

  1. Run the build(s) and/or test(s) that you expect to populate the cache. The first time a new build is run on a particular stack, you can expect no remote cache hits. As part of remote execution, action results are stored in the cache and a subsequent run should pick them up.

  2. Run bazel clean. This command cleans your local cache, which allows you to investigate remote cache hits without the results being masked by local cache hits.

  3. Run the build(s) and test(s) that you are investigating again (on the same machine).

  4. Check the INFO line for cache hit rate. If you see no processes except remote cache hit and internal, then your cache is being correctly populated and accessed. In that case, skip to the next section.

  5. A likely source of discrepancy is something non-hermetic in the build causing the actions to receive different action keys across the two runs. To find those actions, do the following:

    a. Re-run the build(s) or test(s) in question to obtain execution logs:

      bazel clean
      bazel --optional-flags build //your:target --execution_log_compact_file=/tmp/exec1.log
    

    b. Compare the execution logs between the two runs. Ensure that the actions are identical across the two log files. Discrepancies provide a clue about the changes that occurred between the runs. Update your build to eliminate those discrepancies.

    If you are able to resolve the caching problems and now the repeated run produces all cache hits, skip to the next section.

    If your action IDs are identical but there are no cache hits, then something in your configuration is preventing caching. Continue with this section to check for common problems.

  6. Check that all actions in the execution log have cacheable set to true. If cacheable does not appear in the execution log for a give action, that means that the corresponding rule may have a no-cache tag in its definition in the BUILD file. Look at the mnemonic and target_label fields in the execution log to help determine where the action is coming from.

  7. If the actions are identical and cacheable but there are no cache hits, it is possible that your command line includes --noremote_accept_cached which would disable cache lookups for a build.

    If figuring out the actual command line is difficult, use the canonical command line from the Build Event Protocol as follows:

    a. Add --build_event_text_file=/tmp/bep.txt to your Bazel command to get the text version of the log.

    b. Open the text version of the log and search for the structured_command_line message with command_line_label: "canonical". It will list all the options after expansion.

    c. Search for remote_accept_cached and check whether it's set to false.

    d. If remote_accept_cached is false, determine where it is being set to false: either at the command line or in a bazelrc file.

Ensure caching across machines

After cache hits are happening as expected on the same machine, run the same build(s)/test(s) on a different machine. If you suspect that caching is not happening across machines, do the following:

  1. Make a small modification to your build to avoid hitting existing caches.

  2. Run the build on the first machine:

     bazel clean
     bazel ... build ... --execution_log_compact_file=/tmp/exec1.log
    
  3. Run the build on the second machine, ensuring the modification from step 1 is included:

     bazel clean
     bazel ... build ... --execution_log_compact_file=/tmp/exec2.log
    
  4. Compare the execution logs for the two runs. If the logs are not identical, investigate your build configurations for discrepancies as well as properties from the host environment leaking into either of the builds.

Comparing the execution logs

The execution log contains records of actions executed during the build. Each record describes both the inputs (not only files, but also command line arguments, environment variables, etc) and the outputs of the action. Thus, examination of the log can reveal why an action was reexecuted.

The execution log can be produced in one of three formats: compact (--execution_log_compact_file), binary (--execution_log_binary_file) or JSON (--execution_log_json_file). The compact format is recommended, as it produces much smaller files with very little runtime overhead. The following instructions work for any format. You can also convert between them using the //src/tools/execlog:converter tool.

To compare logs for two builds that are not sharing cache hits as expected, do the following:

  1. Get the execution logs from each build and store them as /tmp/exec1.log and /tmp/exec2.log.

  2. Download the Bazel source code and build the //src/tools/execlog:parser tool:

    git clone https://github.com/bazelbuild/bazel.git cd bazel bazel build //src/tools/execlog:parser

  3. Use the //src/tools/execlog:parser tool to convert the logs into a human-readable text format. In this format, the actions in the second log are sorted to match the order in the first log, making a comparison easier.

    bazel-bin/src/tools/execlog/parser \
      --log_path=/tmp/exec1.log \
      --log_path=/tmp/exec2.log \
      --output_path=/tmp/exec1.log.txt \
      --output_path=/tmp/exec2.log.txt
    
  4. Use your favourite text differ to diff /tmp/exec1.log.txt and /tmp/exec2.log.txt.