This is done first and formost to stop "empty" commits that cause errors in
publishDnsUpdates. The reason being that the Cloud DNS api fails when there are
no updates at all in a change.
Allowing this is a requirement for the writer to be idempotent - if we delete a
domain, then run the writer to delete it again - we'll get 0 additions and 0
deletions which fails.
This isn't theoretical either - we've seen it happen, causing a
publishDnsUpdates to fail over and over again.
While fixing this, we also remove all RRS that are common between additions and
deletions. This is just an optimization and shouldn't affect behavior.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=179525218
This is in preparation for running the automatic refactoring script that
will replace all ExpectedExceptions with use of JUnit 4.13's assertThrows/
expectThrows.
Note that I have recorded the callsites of assertions about EppExceptions
being marshallable and will edit those specific assertions back in after
running the automatic refactoring script (which do not understand these).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=178812403
Before pushing an update to Cloud DNS, the CloudDnsWriter needs to read all the domain RRSs from Cloud DNS one by one to know what to delete.
Doing so sequentially results in update times that are too long (approx 200ms per domain, which is 20 seconds per batch of 100) severely limiting our QPS.
This CL uses Concurrent threading to do the Cloud DNS queries in parallel. Unfortunately, my preferred method (Set.parallelStream) doesn't work on App Engine :(
This reduces the per-item time from 200ms to 80ms, which can be further reduced to 50ms if we remove the rate limiter (currently set to 20 per second).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=178126877
Last commit did not pick up all the changes because MOE incorrectly attributed some changes to the wrong commit. This commit should reconcile these. Also picked up some changes to how hamcrest library is depended upon in BUILD file, which should have been included in previous commits.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=177637931
Convert periods to hyphens in multi-part TLDs when using them as a zone name
(cloud-dns doesn't allow periods in zone names).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=172007089
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
The existing CloudDnsWriter code uses ImmutableMap.Builder to construct the
map of DNS records to update. This has been seen to fail on alpha, presumably
in a cases where host records and domain records produce duplicate updates for
a host.
Convert the Builder to a HashMap, allowing us to safely overwrite existing
records in the case of duplicates.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170103421
Right now - if there's an error during DnsWriter.publish*, all the publish from
before that error will be committed, while all the publish after that error
will not.
More than that - in some writers partial publishes can be committed, depending
on implementation.
This defines a new contract that publish* are only committed when .commit is
called. That way any error will simply mean no publish is committed.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=165708063
Attending to this old bug will improve our ability to perform zone comparisons between Datastore and the DNS provider. Right now, zone comparison finds some bogus differences, because the TTL we send to the DNS subsystem doesn't match the TTL we use when generating our local dump files.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=164635557
We're now using java_import_external instead of maven_jar. This allows
us to specify the relationships between jars, thereby allowing us to
eliminate scores of vendor BUILD files that did nothing but re-export
@foo//jar targets, thus addressing the concerns of djhworld on Hacker
News: https://news.ycombinator.com/item?id=12738072
We now have redundant failover mirrors, which is a feature I added to
Bazel 0.4.2 in ed7ced0018
A new standard naming convention is now being used for all Maven repos.
Those names are calculated from the group_artifact name using the
following algorithm that eliminates redundancy:
https://gist.github.com/jart/41bfd977b913c2301627162f1c038e55
The JSR330 dep has been removed from java targets if they also depend
on Dagger, since Dagger always exports JSR330.
Annotation processor dependencies should now be leaner and meaner, by
more appropriately managing what needs to be on the classpath at
runtime. This should trim down the production jar by >1MB. As it stands
currently in the open source world:
- backend_jar_deploy.jar: 50MB
- frontend_jar_deploy.jar: 30MB
- tools_jar_deploy.jar: 45MB
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=143487929
This change replaces all Ref objects in the code with Key objects. These are
stored in datastore as the same object (raw datastore keys), so this is not
a model change.
Our best practices doc says to use Keys not Refs because:
* The .get() method obscures what's actually going on
- Much harder to visually audit the code for datastore loads
- Hard to distinguish Ref<T> get()'s from Optional get()'s and Supplier get()'s
* Implicit ofy().load() offers much less control
- Antipattern for ultimate goal of making Ofy injectable
- Can't control cache use or batch loading without making ofy() explicit anyway
* Serialization behavior is surprising and could be quite dangerous/incorrect
- Can lead to serialization errors. If it actually worked "as intended",
it would lead to a Ref<> on a serialized object being replaced upon
deserialization with a stale copy of the old value, which could potentially
break all kinds of transactional expectations
* Having both Ref<T> and Key<T> introduces extra boilerplate everywhere
- E.g. helper methods all need to have Ref and Key overloads, or you need to
call .key() to get the Key<T> for every Ref<T> you want to pass in
- Creating a Ref<T> is more cumbersome, since it doesn't have all the create()
overloads that Key<T> has, only create(Key<T>) and create(Entity) - no way to
create directly from kind+ID/name, raw Key, websafe key string, etc.
(Note that Refs are treated specially by Objectify's @Load method and Keys are not;
we don't use that feature, but it is the one advantage Refs have over Keys.)
The direct impetus for this change is that I am trying to audit our use of memcache,
and the implicit .get() calls to datastore were making that very hard.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=131965491
