This was a surprisingly involved change. Some of the difficulties included
java.util.Optional purposely not being Serializable (so I had to move a
few Optionals in mapreduce classes to @Nullable) and having to add the Truth
Java8 extension library for assertion support.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=171863777
When trying to run the MapReduce for DeleteOldCommitLogsAction, we run into a
lot of DatastoreTimeoutException during CommitLogManifestReader.next.
This causes the entire shard to fail. Since we have a lot of keys (tens of
millions), this is almost guaranteed to happen, dooming the entire MapReduce.
Here is an attempt to recover from the Timeout Exception by saving the state
before the read, then on failure restoring that state and trying again.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=165172222
Attempting to run DeleteOldCommitLogs in prod resulted in a lot of DatastoreTimeoutException errors. The assumption is that attempting to load so many CommitLogManifests (over 200 million of them), when each one has a slight possibility of failure, has a very high probability of error.
The shard aborts after 20 of these errors, and by eliminating as many loads as possible and retrying the remaining loads inside a transaction we are effectively eliminating any exceptions "leaking" out to the mapreduce framework, which will hopefully keep us bellow 20. At least, that's our best guess currently as to why the mapreduce fails.
EppResources are loaded in the map stage to get the revisions, and CommitLogManifests are only loaded in the reduce stage for sanity check so we don't accidentally delete resources we need in prod. Both of these are wrapped in transactNew to make sure they retry individually.
The only "load" not done inside a transaction is the EppResourceIndex, but there's no getting around that without rewriting the EppResourceInputs.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=164176764
