Concurrency and Thread Safety
As we step towards bigger deployments bringing heavy multi-node and multi-threaded processing, the question of concurrency and thread safety of midPoint code becomes urgent.
Until more elaborate description is prepared, let us provide basic facts/notes here.
Prism and JAXB structures are NOT thread safe
Basic data structures that midPoint works with - prism items and values and JAXB beans - are not thread safe.
So each prism item, prism value and JAXB bean should be owned by a single thread.
There is only single exception to this rule: immutable objects. These can be shared but with caution:
-
Immutability is (still) not enforced consistently. For example, there is no provision for immutability of JAXB beans.
-
When sharing/handing over objects between threads, do not forget about safe publication. This is (still) not addressed properly in midPoint.
DOM trees are NOT thread safe (not even for reading!)
This is really bad.
We have made some measures to deal with this:
-
Appropriate value parsers (PrimitiveValueParser and PrimitiveAttributeParser in DomLexicalProcessor) use synchronization when accessing DOM data.
-
When parsing schemas, we try to use synchronization as well.
Things to fix or consider:
-
Are the above two measures sufficient? (They are probably not especially when dealing with the schemas. They are accessed e.g. when checking for object equality, and probably on some other places. Note that when cloning SchemaDefinitionType we do not clone DOM data. The change was done years ago due to performance reasons and I do hesitate to bring the cloning back.)
-
Is there some other access to DOM data?
-
Generally, it might be worthwhile to replace DOM with some other representation - for primitive values as well as for schemas.
Thread safety of prism definitions is unclear
TODO
OperationResult is NOT thread safe
They are not to be shared across threads. In the case of multi-threaded task executions, each thread uses its own operation result and these are aggregated only after individual threads finish their work. See e.g. https://github.com/Evolveum/midpoint/blob/a76f0b048636ddb916e684a50b984a0f24759a48/repo/repo-common/src/main/java/com/evolveum/midpoint/repo/common/task/AbstractSearchIterativeResultHandler.java#L251-L259.
Tasks are "semi" thread safe
We tried to make tasks thread safe to some extent. But it requires great care on the side of the clients.
A note not directly relevant to thread safety: One should be extra careful when calling methods like incrementProgressAndStoreStatsIfNeeded. They gather statistics from the current thread, so it’s grossly misleading to call them from any other thread (e.g. from the coordinator task thread or from some observer).