Closed tetsuok closed 1 year ago
RFC: @odashi @neubig @pfliu-nlp
Yes it is not negligible amount of degradation that should be fixed.
I also think the current integration tests are basically harmful for rapid development because many of them require to download large amount of data from the Internet. I think we have to:
@tetsuok Thanks. I also have the same feeling and strongly think that we need to re-organize existing integration tests, remove or skip or slim them. (BTW, the reason why some of the existing tests cost too much is that testing training-set dependent features require loading the whole training set. But we probably could add some dummy dataset in DataLab.)
@pfliu-nlp Having heavy tests is not bad if we really need them to achieve better quality, but they should be installed on a separate space to prevent unnecessary burden. I think it may be better to set up some independent place to continuously run these heavy tests.
As a quick solution, I wonder if it'd be possible to set up the heavy tests to run on CI only when we update version.py
?
(EDIT: for now... I agree with @odashi 's solution as a good long-term one)
@neubig I can set up some independent CI using our infrastructure. It is not so difficult to configure I think.
only when we update version.py
Though I thought it is too optimistic, it can also be configured.
(related to this topic, recently I and @tetsuok discussed about versioning of this repository) @tetsuok
Version for a Python package needs to be defined in a single place. (DataLab defines versions in two places)
Correct versioning may be required to follow the newest practice around setup.py
. Created another issue: #490
@tetsuok Maybe you can work with #490 first.
@odashi Sure.
I think this is fixed by #549. @pfliu-nlp and @tetsuok could you confirm and close the issue if so?
Thanks. Verified.
Closing.
commit 8c514c3d81a079d967d208f8bc330c2f202620bb (#437) increases the execution time of
integration_tests.summarization_test.SummarizationTest
. When I measured on my GCP VM, the time of the test increased by 430 seconds (from 6 seconds to 436 seconds), which is too slow to run as automated tests in pull requests. Slow tests need to be removed or replaced with more focused and fast tests. In general, having slow tests leads to productivity drains: Time to update pull requests takes longer, developers would try to include large commits into pull requests to work around slow CI time, pull requests become expensive to review, which makes identifying bugs or design flaws in code review difficult.Repro steps
Output