Attempt retry hook for flaky regression test cases.

[dhermes]

Fixes#531.

To "test" that this works, feel free to add a test case like:

x=0deftest_retry(self):
# Feel free to vary 3 higher and higher, should always be# NUM_RETRIES in the final error message.ifself.x<3:
self.x+=1self.assertEqual(self.x, object()) # Failselse:
self.assertTrue(True)

[coveralls]

Changes Unknown when pulling 09431a8 on dhermes:retry-flaky-regression-cases into * on GoogleCloudPlatform:master*.

[silvolu]

I think that for time based errors (e.g. exceeding QPS for a certain API method) we should bake the retry logic into the library itself, providing default values for max retries, error codes and messages to retry on, but allowing the user to override all of those.
This is partially done in apitools

[dhermes]

@silvolu AFAIK the regression test issues are not due to QPS issues, just intermittent failures or sometimes due to eventual-ness of consistency (though this "shouldn't" happen since usually this is for queries on data inserted weeks prior via populate_datastore.py)

[silvolu]

Oh sorry, got confused with gcloud-ruby (and didn't read this issue with enough attention)

[coveralls]

Changes Unknown when pulling 381e1f9 on dhermes:retry-flaky-regression-cases into * on GoogleCloudPlatform:master*.

[dhermes]

Another one that failed the first time but not on retry:
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/47141980

I took a snapshot of the failure for posterity:

[dhermes]

I didn't think we'd ever see it but two consecutive merges resulted in a 429 Too Many Requests and caused the one I merged to fail. (I somehow lost the screen shot after restarting the build, boo.)

@silvolu Was right that it could occur in our regression tests :)

[dhermes]

@tseaver It seems the last contention here was in allowing AssertionError as one of the retry-able errors. Still having that hangup?

[tseaver]

I'll defer to your report that it failed on the first pass, but passed soon after (although I think we likely have something smelly in the testcase, that would be a different issue).

[dhermes]

@tseaver I'm looking into the smelliness.

The second failure referenced in the bug is due to a non-transactional put() followed by a query. I found 3 or 4 more instances of it in old failed builds.

There are also three tests cases which rely on _generic_test_post which also has a non-transactional put() following by a single-key lookup.

Shall I make these transactional and remove AssertionError from the retry exception list?

---

Also, looking through history it seems this 404 on storage key delete (during module cleanup) occurs pretty often. Our delete code and / or library code seems to have an issue with data staleness.

That also may be smelly and maybe we don't need any retries?

---

The piece I seem to remember failing is in test_query_simple_filter, essentially:

query=datastore.Query(kind='Character', ancestor=datastore.Key('Book', 'GoT'),
[('appearances', '>=', 20)])
expected_matches=6entities=list(query.fetch(limit=7))
len(entities) ==6

but I checked all the failed builds and it's in none of those (it still may have occurred, but in a build we retried).

[dhermes]

For posterity, here are some more "retry"-able errors:

https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/46097250
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/43549205
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/42891293
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/40251811
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/39885452
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/40240783

[coveralls]

Coverage remained the same when pulling 280b7ac on dhermes:retry-flaky-regression-cases into edfd5e2 on GoogleCloudPlatform:master.

[dhermes]

@tseaver I made the datastore put()s transactional and removed AssertionError. I double-checked for funny business in Bucket.delete(). It turns out a new Bucket instance is created and all the objects (Keys) are re-retrieved. So the mystery is still out why we have all these 404's in tearDownModule.

[tseaver]

LGTM

[dhermes]

@tseaver I did some "soul-searching" on this and realized:

1. The datastore flakes are "our fault" in that we didn't have transactional puts. (I'm curious if this ever happens for gcloud-node. Maybe @silvolu knows? I took a look at their last 10 failures and didn't find any evidence.)
2. The retry won't fix tearDownModule, which is the last remaining flake.
3. The storage failure seems to be an API-level issue and manifests in Bucket.delete(). We should loop in storage eng. to see what is expected. (Essentially ls is a lie and it causes delete to fail.)

I'm going to submit a PR (#562) with just the transactional puts and then figure out what to do about the rest.

Adding don't merge label for now.

[dhermes]

@jgeewax Do we have a contact on the storage team to ask questions?

We are seeing a non-trivial number of failures of Connection.delete_bucket(force=True):
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/46346592#L329
https://travis-ci.org/GoogleCloudPlatform/gcloud-python/builds/42891293#L261

---

This failure is because, Connection.delete_bucket():

1. Loops through a list of objects in the bucket.
2. This creates a new _KeyIterator which gets a list of all the objects from the API (since __iter__ does not bind anything to the Bucket object, a new iterator is returned every time

forkeyinbucket:
...

occurs, hence a fresh API request to list the objects)
3. Tries to delete each retrieved key, some of which no longer exist. (hence the 404) This seems to indicate that the '/bucket-name/o/` request is returned stale data. (hence the desire to talk to eng.)

[craigcitro]

summoning @thobrla , expert on all things storage!

[thobrla]

1. Ignoring the 404 is fine. Ultimately, you won't get true atomicity either way (see my response to 4. below)
2. The JSON API does support batch deletes, but it is a performance optimization at best and won't solve this problem.
3. No, there is no API-level support to delete a bucket with objects in it.
4. I'd recommend isolating this to the test code. If you provide this functionality, it must be made extremely clear that this is a non-atomic operation that is not expected to succeed if there are concurrent writers to the bucket. The danger here is providing the illusion of atomicity to the client, especially because that illusion is likely to work at small scale and then fail cryptically at large scale.

[dhermes]

Thanks for the help @thobrla!

I filed #564 to remove the Bucket.delete() method (due to the global namespace of buckets, it's a very serious operation to make so easy to execute). I also submitted PR #563 to address the flaky regression tests.

---

Closing out this now defunct PR. @tseaver I'll keep around the metaclass branch for awhile in case we decide to put it back. Thanks for pushing back about the "something smelly" in our test failures!

I'd guess we would need to address 429 Too Many Requests (so maybe RetryTestsMetaclass will come back) but as @silvolu pointed out it may be best to handle such failures in the library.