Add a script to bisect to a culprit commit. #8442
Conversation
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| # A script to bisect the mainline commits and find the culprit commit. |
There was a problem hiding this comment.
I am trying to understand the difference between this tool and git bisect. According to this comment, it seems that this tool checks only commits on the $BRANCH branch? Also, When I read the code, it seems that this tool also calls ctest repeatedly -- is that for debug unit tests with random behavior? Would it be easier for users to understand about this tool if we have a Github issue listing the design purposes?
There was a problem hiding this comment.
Yes, it only check BRANCH branch. (mainline). git bisect also checks feature branch. I found commits in Paddle's feature branch are not guaranteed to pass tests or compile, so it bisect there won't work anyway.
Yes, the repeat is used to debug flaky behavior. I'll create a issue.
| GOOD = sys.argv[3] | ||
| BAD = sys.argv[4] | ||
| TEST = sys.argv[5] | ||
| BRANCH = 'develop' |
There was a problem hiding this comment.
If we are going to merge this tool, we actually suppose that many users would use it. If so, it seems command line flags would be more convenient than constant definitions.
| # Update PARALLEL for build paralleism. | ||
| # | ||
| # Example: | ||
| # python bisect.py $PWD/../Paddle2 $PWD e80059 6bba41 test_rnn_encoder_decoder |
There was a problem hiding this comment.
Can we remove the first command line parameter (the git repo dir) by looking for the nearest ancestor directory of $PWD which has a .git subdirectory?
There was a problem hiding this comment.
I sometimes create the build directory in a different directory outside the git repo. That way, I can mount an in-memory filesystem for faster build.
|
Thanks, PTAL? |
|
@wangkuiyi It seems I don't have "write access". So I can't merge this PR myself. |
Fix #8449
I expect that we need it to:
Disable test is dangerous. We should first try to find the culprit commit
and fix it or revert it.
I managed to find 3d1ac72 (between 279aa62 and 6773129) that causes
the recent failure of test_rnn_encoder_decoder.