Spring @Transactional support

[koenpunt]

When we first started with spring we used JPA and made heavy use of its relational support. This caused some issues with spring-graphql because sometimes lazy associations were accessed in a field resolver, causing an error that said something like "there isn't an active transaction".
We worked around this by writing a custom instrumentation to start a transaction for every query and mutation.

Now that's we removed all associations (in preparation of migrating to R2DBC), we no longer need this transaction by default. But removing this instrumentation somehow causes a significant increase in queries.

Adding @Transactional to graphql controllers doesn't seem to make a difference. Also adding @Transactional to a parent controller, and @Transactional(propagation = Propagation.MANDATORY) to controller of an underlying type controller, causes the execution to fail with:

org.springframework.transaction.IllegalTransactionStateException: No existing transaction found for transaction marked with propagation 'mandatory'

This seems odd to me, because I would expect a transaction to exist there because of the transaction started in the parent controller. After some discussion with a teammate I've realized this isn't odd, because both controllers are invoked separately by the datafetchers, and thus not sharing the transaction. It still would be very welcome to have a solution for this.

Is there any support for transactions in Spring Graphql? Or is using a custom instrumentation (or ExecutionStrategy) the only way out for this?

[nort3x]

I'm not 100% sure but i think jpa.open-in-view=true can solve this,
I know it's not recommend and I'm waiting for an answer on this as well

[mp911de]

Spring GraphQL uses a DataFetcher-oriented execution approach in which nested fields are retrieved by subsequent calls to nested DataFetchers once the parent has been materialized.

Once a DataFetcher call has been completed, the GraphQL engine introspects the returned object and uses a collection of DataFetchers to resolve potentially nested properties.

Because of this arrangement, it is impossible to declare a global transaction via e.g. @Transactional as there is no single entry point, contrary to Spring WebMVC where all activity happens within a controller method.

You see an increased number of queries because that's what happens here. Looking up references by their identifier is recommended to avoid lazy-loading issues. However, that approach comes at the cost of additional queries.

JPA implementations typically optimize to some extent using JOIN queries or batch lookups to avoid database roundtrips. They can leverage caches to avoid database hits if the objects are already in the session/2nd level cache.

You could de-proxy any objects within your DataFetcher or map these onto a DTO (another form of de-proxying). Still, such an approach requires a mapping utility such as MapStruct or manual copying of properties.

Spring GraphQL and the GraphQL engine can leverage reactive programming and asynchronous execution patterns, making it impossible to have ThreadLocal transactions out of the box as we cannot make any assumptions over the actual thread on which a call is being made.

[koenpunt]

[...] to have ThreadLocal transactions out of the box as we cannot make any assumptions over the actual thread on which a call is being made.

Not that I think that would be a solution on its own, but aren't there database clients for spring that support transactions over reactive streams?

Also; thanks for the extensive response.

[mp911de]

There is reactive transaction support, but since there is no @Transactional entry-point for the entire GraphQL request processing, we cannot start/clean up the transaction.

[koenpunt]

but since there is no @Transactional entry-point for the entire GraphQL request

Isn't the entrypoint where the @QueryMapping or @MutationMapping annotation is?

[mp911de]

No, the entry point is the WebGraphQlHandler that holds on to DataFetchers (typically the one that handles the Query element in the request document) and interceptors. The config infrastructure creates DataFetchers for annotated methods.

[rstoyanchev]

A more concrete sample or snippets of schema types, entity objects, and controller methods would help.

If it is only a @QueryMapping method then yes, that is a single point of entry. However, often there are additional @SchemaMapping methods for nested fields, and no single point of entry, at least not in application code. The last such point is the web interception chain that calls ExecutionGraphQlService and that in turn calls GraphQL Java which invokes DataFetchers, and that's what controller methods are registered as.

[rstoyanchev]

Keep in mind that even then, there is no single Reactor chain since GraphQL's AsyncExecutionStrategy invokes DataFetchers and expects CompletableFuture. We adapt Flux and Mono to that but it means it's not a fully connected Reactor chain through which context can propagate from start to end. Only a number of independent smaller chains, one for each DataFetcher. We do however propagate context from the transport layer to each of those smaller chains, so they all see whatever comes from that level.

[xenoterracide]

If the data fetchers are tied to the request, and the request received something like multiple mutations or even nested queries wouldn't it be possible to wrap the entire request in a transaction?

[rstoyanchev]

We've had an internal discussion.

The recent upgrade of our context propagation in #459 provides some new capabilities that make it easier to propagate context across data fetchers, and not only from the transport level to data fetchers. What we need to investigate are some options to hook into the start and end of a transaction.

For example if @Transactional is on a @QueryMapping method, then we need to start a transaction before this is invoked, and then close it after the entire query processing is done, which is obviously not the same as when the @QueryMapping method exits. That would be a place to start.

A single @MutationMapping method is easy to surround with a transaction. I don't know if nested mutations are possible or an option in GraphQL Java? For multiple mutations, there isn't anything to put @Transactional that would surround all of them, and so the question there is how to indicate transaction boundaries, perhaps a schema directive.

For @Transactional on a nested @SchemaMapping, it seems relatively straight forward in terms of participating in an existing transaction (assuming we start propagating transactional context). It is less clear what to do if the intent is to start a transaction, as it's not clear how we would hook into the right place for when it should end, or perhaps we might simply not support that.