This completes the exhaustive refactoring of foundational temporal types from Joda-Time to the native java.time API across the entire codebase.
- Replaced org.joda.time.DateTime, Instant, LocalDate, and Duration with java.time equivalents.
- Audited and updated Clock implementations (FakeClock, SystemClock). Added nowMillis(), nowDate(), and nowDateTime() to eliminate repetitive conversions and maintain parallel naming.
- Replaced ZonedDateTime with OffsetDateTime globally per go/avoid-zdt. OffsetDateTime is a better fit as we use a hardcoded ZoneOffset.UTC throughout the system, making geographical time zone rules (like daylight saving time) irrelevant and preventing serialization ambiguities. Added a presubmit check.
- Completely removed all transitional bridge methods from DateTimeUtils and deleted obsolete converters (e.g., DateTimeConverter).
- Updated testing infrastructure, Apache Beam pipelines, custom JCommander parameters, and networking modules to solely rely on java.time primitives.
- Retained the lone necessary org.joda.time.Instant usage in SafeBrowsingTransforms required by the Apache Beam API.
- Cleared Gradle lockfiles and removed the joda-time dependency entirely from the build configuration.
The fix for https://hibernate.atlassian.net/browse/HHH-20276 was
included in 7.3.4 and will be included in version 8 in the future. 8.x
is still in alpha though so we don't want to use it yet.
We want to make sure that we use the same XML factories no matter what,
so we use "newDefaultFactory" instead of "newFactory" (to avoid picking
up some random thing on the classpath).
This also fixes an exception that occurs if you haven't synced the
internal repo with the public repo.
Error-prone introduced many more checks in Java 25. We fixed a few
and suppressed most. A follow-up bug is opened to clean this up.
An ai agent should be able to clean up most of it.
This PR is created with gemini-cli. Summary of experience:
* The good: AI caught most compatibility issues, and with permission,
suppressed them through compiler flags and errorprone options.
It also caught many versio references in scripts.
* Where it didn't shine:
- It did not find and update the target version spec in the custome
VKey annotation processor source file.
- It did not flag eclipse-temurin:21 docker image for upgrade.
- When running into failure, its first instinct is to disable checks
e.g., -Werror instead of fixing them.
* More Gradle 9 preparations
Fix additional compatibility warnings after upgrading to Gradle 8.14.3
from 8.13.
* More Gradle 9 compatibility fix
More fixes after upgrading Gradle from 8.13 to 8.14.3.
Upgraded the gradle-license-report plugin, and handled config leaking
issues.
Researched using gemini web and manually applied the fixes. Gemini-cli
could not find the right solution.
This ended up being wayyyy more complicated than expected due to
issues with Hibernate, various dependencies having conflicts with the
proto dependency version, and other breaking changes.
Notes:
- Hibernate 7 switches up the user type / converter system and for us,
this means we must be / want to be more explicit with how we convert
and store things. For example, we need to add Postgres types to @Column
definitions.
- Hibernate 7.3 has an issue with generic MappedSuperclasses -- we have
issues with BaseDomainLabelList. I'll investigate that, but for now
let's stick with 7.2.x
- H7 is more strict with annotations and prevents us from storing mapped
superclasses embedded within other objects. This kinda makes sense but
makes the History objects a bit more difficult. We had to add "concrete"
embeddable DomainBase and HostBase objects that we can store/retrieve
from the DB.
- We convert some of the calls to "Query" to "TypedQuery" -- in
Hibernate 8 / JPA 4.0 these will be super-deprecated and we'll need to
shift everything over, so this is necessary.
- You aren't supposed to put callback listeners on embedded entities
(because it can be not obvious what's happening). We don't like that,
so we add our own annotations that are processed recursively for
embedded entities, so we get things like the update / create
timestamps.
- Hibernate doesn't allow for multiple converters to be auto-applied to
the same "type" and it counts all VKey converters as one type.
Unfortunately, this means we have to explicitly mark each one.
- A bunch of other dependency changes were required to keep from having
the proto 3/4 conflict
Ben McIlwainandGitHub
gbrodmanandGitHub
Weimin YuandGitHub
Lai JiangandGitHub