* Compare migration data with SQL as primary DB
Add a BEAM pipeline that compares the secondary Datastore against SQL.
This is a dumb pipeline to be launched by a driver (in a followup PR).
Manually tested pipeline in sandbox.
Also updated the ValidateSqlPipeline and the snapshot finder class so
that an appropriate Datastore export is found (one that ends before the
replay checkpoint value).
We can use it more places later but this can serve as a template. We
should inject the connection to the read-only replica (only created
once) to the constructor of the action, then use that instead of the
regular transaction manager.
We add a transaction manager that simulates the read-only-replica
behavior for testing purposes as well.
In addition, we set the transaction isolation level to READ COMMITTED
for this transaction manager (this is fine since we're never writing to
it). Postgres requires this for replica SQL access (it fails if we try
to use SERIALIZABLE) transactions. We didn't see this with the pipelines
before since those already had transaction isolation level overrides
* CommitLog handling code should call ofyTm
The tm() call will use JPA transaction manager after the switch-over to
SQL. These calls would lose their transaction semantics.
Both actions are to be invoked after the switchover in case we have to
switch back to Datastore as primary.
* Only compare recent changes in Datastore and SQL
When comparing Datastore and SQL, ignore older History and EPP resource
objects. This cuts the run time in half compared with a full comparison.
The intention is to run a full comparison before the switch-over from
Datastore and SQL, and run this incremental comparison during the down
time.
The incremental comparison takes about 25 minutes in production.
Performance can be improved further by filtering out older billing
events (OneTime and Cancellation). However, we don't think further
optimization is worth the effort (considering that Recurring events
cannot be filtered since they are mutable but without lastUpdateTime).
Verified in Sandbox and prod with and without time filter.
We already have ValidateEscrowDepositCommand to check for internal
reference consistency of two deposits, i. e. making sure that all
contacts and hosts referenced by domains exist in the same deposit.
Therefore to compare whether two deposits are equal we only need to make
sure that they contain the same domains and registrars, assuming they
both pass the validation. We don't compare their contents directly
because the MapReduce deposit contains all contacts and domains whereas
the Beam deposit only contains referenced ones, making a direct
comparison impossible.
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/google/nomulus/1476)
<!-- Reviewable:end -->
* Speed up updating of premium lists
There are two parts to this:
1. Don't load the premium entries in the command prompt (this isn't
necessary and we didn't display that information anyway).
2. Set a proper batch size (rather than just 1) when saving all the
premium entries. This means that we generate only one INSERT statement
rather than N statements.
* Allow usage of a read-only Postgres replica
This adds the Dagger provider code for both the regular and the BEAM
environments, which are similar but not quite the same.
In addition, this demonstrates usage of the replica DB in the
RdePipeline. I tested this on alpha with a modified version of the
RdePipeline that attempts to write some dummy values to the database and
it failed with the expected message that one cannot write to a replica.
* Resolve ResaveEntityAction related conflicts
* Replace string with existing constants
* Remove solved TODOs related to ofy string to new vkey string
* Add a TODO for clean up
* Fix missing annotation
* Make not logged in errors take precedence over extension errors
This is the right order to do the checks in, because if the registrar isn't
logged in (or their login failed) then they will have an empty set of declared
extensions, so any attempt to use an extension will throw a "Service
extension(s) must be declared at login" error. This is potentially misleading
because the actual error in this situation is that the registrar isn't logged
in at all.
This also fixes some flows that weren't declared final (but should be), or
methods declared final on final classes, which is superfluous.
* Don't throw errors when existing premium list is empty
This state is possible to get into when things go wrong and it shouldn't prevent
saving new revisions of the list. Note that it will continue to throw errors if
you attempt to save a new revision that is blank (which is usually a mistake).
See http://b/211774375
* Make premium list saving run as a single transaction
This fixes the bug where the new revision is saved, but then execution gets
halted for some reason (e.g. request timeout) before the entries finish saving,
which leaves the DB in a bad state with a new top revision containing zero
entries, thus making everything standard.
* Add pending action extension to server update poll messages
This is necessary for the poll messages to contain the necessary context
explaining what domain name the relevant statuses were being added/removed
to/from.
* Always use JPA TM on Beam
Beam does not have access to datastore. Using ofy on Beam always results
in an error. Normally we should use database migration state schedule to
determine which TM to use, but on Beam there's no point in doing so. By
hard-coding the TM on beam to be SQL we can start testing features before
we migrate to SQL mode, for example the new RDE pipeline.
Also made a change to where the manual deposits are stored. It made more
sense to store them under manual/[direcitory]/[jobname]/ instead of
[jobname]/manual/[directory]/.
TESTED=deployed the pipeline on production and ran a job.
This version of Beam does not have an explicit dependency on log4j.
There are a couple of other things that need to change due to the
upgrade.
1) The new version pulls in a dependency that is not on Maven Central
but on packages.confluent.io, so we need to explicitly add this repo.
2) The new version has a dependency on flogger 0.6 anb above , which removed
the LoggerConfig class (see google/flogger#142).
We therefore backported the class. In the long term we should do what
was suggested in the issue and use the normal JDK Logger config
directly.
3) The intSqlPipeline dependency graph also needs to be updated.
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/google/nomulus/1472)
<!-- Reviewable:end -->
* Ignore Prober related entities when comparing db
Deletion of prober entities are not propagated to SQL, resulting in two
types of mismatches: entity only exists in SQL, or copies of an entity
differ in deleteTime. Both cases should not count as erros.
* Make ImmutableObject.toString deterministic
Remove the identity hash from the output. There is no use case
(including debugging) for it.
Removing it allows us to also remove some overriding implementations in
subclasses, and may also simplify tests.
* Improve logging/comments for commit log forks
It looks like the diff file lister is doing a second constructDiffSequence()
when a commit diff file is missing from the final sequence for purely
informational purposes. However, this purpose wasn't clear when investigating
an actual case of this.
This PR adds another warning to hopefully make the log output a bit more
useful, and also promotes the "gap" log message to a warning and adds a
comment indicating the purpose of the second constructDiffSequence().
This adds two new options:
1) An option to run RDE in lenient mode.
2) An option to run RDE with the new Beam pipeline regardless of the datastore setting.
<!-- Reviewable:start -->
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/google/nomulus/1453)
<!-- Reviewable:end -->
* Copy into PersistentSets in Domains if applicable
This is similar to https://github.com/google/nomulus/pull/1456
It is possible that in some cases we could get an exception:
Caused by: org.hibernate.HibernateException: A collection with cascade="all-delete-orphan" was no longer referenced by the owning entity instance: [parent]
The main cause of this, according to research (StackOverflow :P) is that
when Hibernate is calling the setters for these sets of children it's
losing the connection to the previously-managed child entity (which it
needs, in order to know how to delete orphans). Thus, the solution is to
maintain the same instance of the persistent set and just add/remove
to/from it as necessary.
This is complicated by the fact that sometimes the setter is given the
persistent set (the one we want to keep) and sometimes (?) it isn't.
In replay (and possibly in other cases) we're getting an exception:
Caused by: org.hibernate.HibernateException: A collection with cascade="all-delete-orphan" was no longer referenced by the owning entity instance: google.registry.model.domain.DomainHistory.internalDomainTransactionRecords
The main cause of this, according to research (StackOverflow :P) is that
when Hibernate is calling the setters for these sets of children it's
losing the connection to the previously-managed child entity (which it
needs, in order to know how to delete orphans). Thus, the solution is to
maintain the same instance of the persistent set and just add/remove
to/from it as necessary.
This is complicated by the fact that sometimes the setter is given the
persistent set (the one we want to keep) and sometimes (?) it isn't. We
will need to try this out to be sure.
The cardinality for the paths is unbound, and could generate a huge
amount of metrics if someone is scanning our web WHOIS endpoint.
See b/209488119 for an example of such a sudden increase in metric volume.
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/google/nomulus/1451)
<!-- Reviewable:end -->
* Add replicateToDatastore to non-prod cron files
This shouldn't do anything yet (since ReplicateToDatastoreAction checks the
migration state before doing anything) but we'll want to have this in
place.
* Filter out empty dsData objects, not just null ones
Hibernate/SQL will get mad if the digest is null or empty, and
previously we only check for null. We should filter out empty digests as
well.
* Properly handle Joda Money in JPA
Joda Money has BigDecimal as amount, which is mapped to a numeric(19,2)
column in the database. As a result, the Money amount load from DB has
scale 2. This becomes a problem with currencies such as JPY, which
requires scale to be 0. To properly load a currency, we must adjust the
scale post-load.
The current approach, which uses Hibernate component mapping, puts the
burden of post-load cleanup on each entity type that uses Money. It is
easy to forget this, as we just discovered.
This PR uses a CompositeUserType to map Money. It adjusts the scale
properly when loading Money instances. Although CompositeUserType appear
to be deprecated in Hibernate 6, it is the only proper solution right
now for mapping non-owned classes.
This is what's causing https://b.corp.google.com/issues/208274109, where
there are DTR rows with null foreign key values.
We should probably wait to make the columns officially non-null until we
get this in and verify that we can do so.
* Write commit logs during SQL->DS replay
Previously, we had no way to ignore read-only mode while still writing
commit log backups. Now, we added this so we can write commit logs in
the SQL->DS replay.
Note:
- When moving to either of the DATASTORE_PRIMARY stages, one must
manually set the SqlReplayCheckpoint first. We don't write to SQL with
backup in this stage because we already wrote the transaction in
question to Datastore. The fact that we manually set the replay
checkpoint means that we'll ignore the extra commit logs that might
otherwise cause problems if we switched back and forth from
DATASTORE_PRIMARY to SQL_PRIMARY.
- The commit logs written during the SQL_PRIMARY phase will, ideally, be
unused. We write them here only so that in the event of a rollback to
Datastore, we will have them for RDE purposes.
This is a result of bad data (we should never allow a null digest) and
we'll need to fix that separately, but this allows us to not fail on
this during replay
* Add NotLoggedInException tests to flows and flow docs
This wasn't included in flows.md before because the test existed in
ResourceFlowTestCase. So even though the exception could be thrown and
even though this was tested, it wasn't picked up in the documentation
because the documentation is picked up from the corresponding concrete
test class.
* Validate SQL with Datastore being primary
Validates the data asynchronously replicated from Datastore to SQL.
This is a short term tool optimized for the current production database.
Tested in production.
We want to keep the read-only-mode-exception as an unchecked exception,
so we introduce a temporary check in the EppController that provides a
specific error message for this situation (rather than letting it fall
through to the generic "command failed" messaging
* Replace with stringify() and VKey.create(string)
* Convert implicit cases of VKey.fromWebsafeKey(string)
* Convert from Key to VKey to use stringify()
* Modify existing code to show correct string representation of a key
* Use VKey.create(websafeKey) to get ofy key in ResaveEntitiesCommand
* Add TODO note in CommitLogMutation and determine if key string should be modified
* Revert from stringify() to getOfyKey().getString()
* Add bug ids to TODOs
* Ignore read-only mode in SQL->DS replication process
We need to be able to save indices and save data about the replication
even when we're in read-only mode.
