GraphQL Performance Issues (using Absinthe)

One key thing to check on is what your scalars are doing. Scalars are serialized in the result phase, which happens entirely after the resolution phase. In the past, poorly optimized serialization functions have lead to lengthy post-resolution times.

To dig in further, some additional metrics would be handy. Can you provide an example query document? How large is the result JSON in bytes? How large is the schema (approximated by MyAppWeb.Schema.__absinthe_types__ |> map_size)?

Thanks for responding, @benwilson512!

Here’s a gist for a sample GQL doc that would generate the flame graphs in my original post. I’ve taken some fields out for IP purposes, but that’s generally the shape. Most types you can assume the type of. There are aggregate types when it makes since (user, for example). Also of note, we’re not currently using any custom serialization.

The result of __absinthe_types__ |> map_size is 34, which doesn’t seem to be that large to me, but perhaps that’s incorrect?

Update
So I’ve shaved some time off because I wasn’t .uniq-ing IDs in batch loaders, but there are still two sizable time gaps for which we’re unable to account:

1. after context is added, before resolution
2. after the resolver is run.

Hi @derekbrown I still need some sense of the size of the returned document. This will help diagnose time spent after the resolution phase constructing the actual JSON result.

While

I recommend using Telemetry to provide a few more data points, specifically the [:operation, :start] event would be very helpful because right now although it is likely that Absinthe starts work immediately after the context is set, in theory there could be other plugs or issues within Absinthe.Plug that are introducing delay. See Telemetry — absinthe v1.11.0 for details.

The document is relatively small, so processing on the document should happen pretty quickly. I’m going to take that document and create a project with a similar schema to see if we can reproduce the observed performance issues.

Sorry about that! Missed the question in your first reply, @benwilson512 . The return document size ranges depending on the query variable, obviously, but the return document for the graphs above was 155kb. FWIW, we see the same “lag” regardless of the size of query.

We have Telemetry events already setup; I can get that to you ASAP. Not a bad idea to add those events to the tracer as well. If I’m able to do that easily, I’ll update this thread with new graphs that show the start & stop events on the operation.

As a side note, it will be super interesting to see this data alongside a PersistentTerm schema once we’re able to release that.

Update: I’ve been able to add Absinthe Telemetry events to the tracer, per reply above from @benwilson512 . Below are some observations and pretty pictures. I’ve tried to keep the same query here from the original post, for some apples-to-apples comparison. Hope this all helps.

• You’ll notice there’s a significant reduction in speed. The query document remains the same, along with the response size, more or less (if anything, the response is likely bigger). Part of this is attributed to some logic/DB work we’ve been doing underneath the query, and some of it is directly attributed to the .uniq issue I mentioned earlier. Hope that helps anyone else with looking to increase performance.

Overall shape of the query. The first flame of spans is the context loading, followed by the operation.start event, followed by the first resolution.start event. We’ll zoom in subsequent screenshots, but there’s a significant gap between those two events, and after the operation.stop event.

Screenshot 2020-05-25 12.35.392282×482 39.6 KB

Context to Operation Start. There’s a miniscule gap (≈300µs) between context being added and the operation starting, but not significant enough for us to blink.

Screenshot 2020-05-25 12.36.48880×396 30 KB

Operation Start. The span on the right of the screenshot is the first resolution.start event, which is followed by the fetch_messages query in the overall screenshot above. Would we expect there to be a ≈6ms delay or so between operation start and resolution start?

Screenshot 2020-05-25 12.36.323248×358 42.3 KB

Operation Stop. There’s a ≈10ms gap after operation.stop that’s unaccounted for. I’ll dig into what could be happening there (perhaps other plugs?). I don’t assume so there’s any cleanup or anything that Absinthe would be doing after the operation, is there?

Screenshot 2020-05-25 12.37.193292×410 53.1 KB

Edits

1. Maybe a clue, or could be a rabbit trail: it seems that the delay to context matters whether or not it’s a query or mutation. The gap from trace start to context being added is much less for mutations.
2. The above is a rabbit trail. Doesn’t seem to be anything deterministic causing the delay before Context gets called.

Running into dead ends here trying to measure what’s happening before context gets added, or after the resolver responds with the data.

I’ve tried removing Plugs, slimming down the endpoint/router logic, and adding spans to field-specific resolvers, all to no avail.

Open to other suggestions; I just want to know what’s happening.

The time spent between the operation starting and the first resolver running is time spent parsing and validating the document. The time spent between the last resolver and operation stop is time spent turning the raw Elixir results from resolution into the nested string keyed map that will ultimately end up as JSON, and validating / serializing scalars. Time spent after the operation has stopped is gonna be mostly JSON encoding I think.

With respect to the comment you made in the github issue, the resolution phase is expected to be the longest because it runs the resolution functions. The sum of the time spent in the resolution functions will be less than the time spent in the resolution phase due to the overhead of walking the tree and its associated results.

Thanks, Ben! Makes sense. You’ve been super helpful and I’m appreciative.

And JSON encoding being a majority of the tail makes sense as well; I think that length of time is proportionate to our response object. I’m still trying to diagnose if it’s our nested query documents (sometimes 4-5 levels) that’re causing the perf gap during validation or if it’s something else.