[keycloak-dev] Keycloak with CockroachDB

[Poiffaut Romain]

Hello,

For the Cloudtrust project (https://github.com/cloudtrust), I have investigated the feasibility of supporting CockroachDB in Keycloak.

I have created a fork based on Keycloak 3.4.3 (it was the latest version when I started on this) working with CockroachDB 2.0.2. It is now working with this DB, and passes all unit tests. Some resilience tests have also been performed to validate the whole integration.
Address of the repo : https://github.com/cloudtrust/keycloak/tree/cockroach-db

For those who might be interested, here are some details:

CockroachDB is a multi-master SQL database designed to run in the cloud and being resilient to failures (https://www.cockroachlabs.com/).
This database has a lot of very interesting properties such as being lockless, distributed and supporting serializable isolation. 
CockroachDB introduces the notion of SAVEPOINT. As this DB is lockless, a transaction may fail due to a concurrent transaction. In such case, we can rollback to the SAVEPOINT and retry the transaction. Retrying transactions has the benefit of increasing their priority each time they are retried, thus increasing their likelihood to succeed. (More detailed information are available in their very good documentation (e.g. https://www.cockroachlabs.com/docs/stable/transactions.html#client-side-transaction-retries, https://www.cockroachlabs.com/blog/how-cockroachdb-distributes-atomic-transactions/, https://www.cockroachlabs.com/blog/serializable-lockless-distributed-isolation-cockroachdb/ ))

So even if CockroachDB uses PostgreSQL driver to communicate with the DB, one of the challenges was to add an automatic transaction retry mechanism with the smallest impact on Keycloak.
Thanks to the architecture of Keycloak, this mechanism can be added in KeycloakSessionServletFilter with a very limited impact. 

A second challenge is due to the rollbackOnly mechanism implemented in Keycloak and Hibernate: after a rollback, a transaction cannot be used anymore.
The retry operation must be performed in the same transaction to increase its priority.
Thus the rollbackOnly mechanism is disabled/bypassed in order to keep the transaction active even after a rollback is issued.
As suggested by CockroachDB, we replace the default Hibernate transaction coordinator class to a custom one (https://github.com/cockroachdb/hibernate-savepoint-fix).
Moreover, we mainly modify JpaKeycloakTransaction so that if the transaction fails to commit due to retryable transaction error, we disable the rollbackOnly mechansim to able to retry the transaction.

CockroachDB does not support addition of some constraints (e.g. primary keys) after table creation.
To circumvent this limitation, we can create a new table, migrate the data, delete the old table, rename the new table with the correct name.
As CockroachDb was not supported by Keycloak until now, we didn’t adapt all existing liquibase scripts. We decided to create a new liquibase script which creates the whole database schema for the current version.
This current limitation is being discussed and will be fixed in future release (https://github.com/cockroachdb/cockroach/issues/19141).

Some tests have also been slightly adapted to support SERIALIZABLE isolation, so other DBs configured with such level can also benefit from this adaptation (i.e. PostgreSQL)

These challenges have been solved and our forked version of Keycloak is now compatible with CockroachDB, but it currently is at the cost of breaking the usage of standard databases.
Our company is really interested to add the support of this DB into Keycloak and to provide it to the community.
The next steps now would be to migrate our fork to the latest version of Keycloak and add the support of this DB without breaking support of the others and we would be happy to discuss it. 


Cheers,
Romain Poiffaut