diff options
author | 2018-05-23 07:22:06 -0700 | |
---|---|---|
committer | 2018-05-23 07:23:10 -0700 | |
commit | 24be509e3a40a0ac661edaf825f6dcd5d9a94d47 (patch) | |
tree | 2572a605670ebce2a61346af2f107db96d674a61 /src/main/java/com/google/devtools/build | |
parent | 17735dd9b0bbde43d73a6539733b02b07bec6abc (diff) |
Automated rollback of commit 724bdbfa47576c67eeec5c74d594203fe05188c7.
*** Reason for rollback ***
Change introduced flakiness in //src/test/java/com/google/devtools/build/lib:actions_test
Verified the culprit commit using:
bazel test //src/test/java/com/google/devtools/build/lib:actions_test --runs_per_test=100 --notest_keep_going -t-
See https://github.com/bazelbuild/bazel/issues/5242.
*** Original change description ***
Relax the threshold for calling getDigestInExclusiveMode().
If the policy goal is to minimize disk seeks, this should be equally good.
RELNOTES: None.
PiperOrigin-RevId: 197720793
Diffstat (limited to 'src/main/java/com/google/devtools/build')
-rw-r--r-- | src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java | 13 |
1 files changed, 7 insertions, 6 deletions
diff --git a/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java b/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java index 32dbbcea2b..d16d1bae89 100644 --- a/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java +++ b/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java @@ -237,12 +237,13 @@ public class DigestUtils { // All right, we have neither a fast nor a cached digest. Let's go through the costly process of // computing it from the file contents. - if (fileSize > 128 * 1024 && !MULTI_THREADED_DIGEST.get()) { - // We'll have to read file content in order to calculate the digest. - // We avoid overlapping this process for multiple large files, as - // seeking back and forth between them will result in an overall loss of - // throughput. Files shorter than the default readahead window of 128KB - // on Linux are assumed to be readable in one seek. + if (fileSize > 4096 && !MULTI_THREADED_DIGEST.get()) { + // We'll have to read file content in order to calculate the digest. In that case + // it would be beneficial to serialize those calculations since there is a high + // probability that MD5 will be requested for multiple output files simultaneously. + // Exception is made for small (<=4K) files since they will not likely to introduce + // significant delays (at worst they will result in two extra disk seeks by + // interrupting other reads). digest = getDigestInExclusiveMode(path); } else { digest = getDigestInternal(path); |