async_hooks experimental status

[martinkuba]

We are looking at using async_hooks to maintain state in our instrumentations when async/await is used. What is your recommendation for using async_hooks in production code at this point?

We realize that it is marked as experimental API. However, it is not gated with a flag. Is the API likely to change? When do you think it will come out of experimental status?

[tniessen]

cc @nodejs/async_hooks

[refack]

@martinkuba hello, a little bit of context:

1. The module isn't gated with a flag since the code that implements it is very pervasive, and essentially it's impossible to turn it off, so a flag would have been a "false flag".
2. Indeed the experimental demarcation is there to indicate that the API might change with no correlation to semver rules. But the work that has been done during the v8.x cycle tried to be semver-minor nonetheless (refactored APIs were back-shimmed for compatibility, e.g. async_hooks: rename currentId to currentAsyncId #13490).
   The implementation itself is not considered buggy in anyway.
3. Some big players are betting on async_hooks for tracing and APM, but I don't know first hand of any big projects that use it in production, maybe the others can help?

My very subjective estimation is that the "user" API will not change much. Most of the work being done is on the "embedder" API i.e. code that wants to trigger the hooks, not code that wants to use the hooks and query the graph.

IMHO go for it, but I'm a bleeding edge kind of guy, so it is very dependant on your needs and organizational constraints.

[AndreasMadsen]

In my opinion: treat it as if it were behind a flag.

I can almost guarantee there are going to be API changes, although we will try to make them backward compatible in 8.x, but that is only until it becomes to inconvenient.

During 8.x we have already deprecated 6 function calls or something like that.

[martinkuba]

@refack @AndreasMadsen Thanks for your replies, I really appreciate it.

Have there been any performance tests done - to determine overhead and also to look for memory leaks when running over long period of time? If yes, can you please point me to those tests?

We are running benchmark tests ourselves, and would like to compare with any existing results.

[refack]

@martinkuba that's a good questions re:

• https://github.com/nodejs/benchmarking
• https://www.youtube.com/watch?v=r0tdN5gx5eU&t=26s
• https://www.youtube.com/watch?v=F0rTqMezYVo
  It is still an open question how should node core do benchmarking...

In this case I can only contribute anecdotal data-points. Personally I know of one project (small-medium) that uses async-hooks in a long running system in production, they are not doing rigorous benchmarking, but have not been noticing memory leaks or performance degradation. It's also a complicated data-point to analyse since they upgraded from 6.x specifically so they could use async-hooks...

1. I believe everyone would be happy to hear New Relic's results, if you could share them.
2. I'll post a Request For Comments from the general public.

P.S. https://twitter.com/refack/status/896460320079785987 - #14794

[AndreasMadsen]

@mcollina asked about the plans for the experimental status. Let's have the discussion here.

What I would like to see done before moving out of experimental status:

• Setting up the benchmark suite to be run with and without async_hooks and see that there is only a small measurable difference given our tolerance.
• Deprecate runInAsyncIdScope as agreed in Remove async_hooks runInAsyncIdScope API #14328.
• Finish discussion on setTriggerId in Remove async_hooks setTrigger API #14238. I think we need others opinion on this, the discussion between me and Trevor is not very productive.
• Finish discussion on newUid, emitInit, emitBefore, emitAfter and emitDestroy in Remove async_hooks Sensitive/low-level Embedder API #15572. I still think these are unnecessary and prone to cause errors and invalid states.
• Once the three API discussions have been completed, we should separate the internal API from the external API. Similar to what I proposed in https://github.com/AndreasMadsen/node/pull/1.
• Opt-in async_hooks state validation. We had decent state validation before (async_hooks,async_wrap: fixups and fewer aborts #14722 and async_hooks: triggerAsyncId can be undefined #14387) but it was always active which was very problematic. PR: async_hooks: skip sanity checks when disabled #15454

edited: added new links to #15572 and #15454.
edited: spelling and poor phrasing.

[trevnorris]

@AndreasMadsen

Opt-in async_hooks state validation. We had decent state validation before (#14722 and #14387) but it was always active which was very problematic.

Oh, good idea. I was thinking of just leaving it for debug build, but adding a flag or env var is an excellent idea. You have any ideas about how it should be implemented?

And for future generations reading this: the only reason I removed some of the validation was because of the ecosystem breakage and not being able to find an alternative solution quickly enough. If the edge cases that caused those failures could be determined I'd like to eventually re-add all state validation.

[AndreasMadsen]

@trevnorris Well, as I have mentioned a few times I think we should just check async_hook_fields [kTotals] > 0. I realize this doesn't ensure correctness if only executionAsyncId() or triggerAsyncId() is used, but it is so trivial to add some empty hooks and we can document that "hack". One could argue that async_hooks is useful with incorrect states, I honestly think it will only lead to confusion and fewer critical bug reports. If it turns out to be a problem, we could change the behavior when moving to a stable state.

[refack]

Just for context, AFAIK there hasn't been a critical async_hooks in about 2 months (I think #14387 was the latest), so the code is converging anyway.

[AndreasMadsen]

@refack That is likely a false positive, there hasn't been an issue because we removed/relaxed the asserts. That doesn't mean there is no async_hooks state errors. It just means we are not seeing them anymore.

[mcollina]

@AndreasMadsen we cannot ship a Node.js release with tight asserts by default. If we need them to exists, then let's put them behind a flag. In this way, we can test this in applications without causing problems for everybody else that might not even be using async_hooks.

[addaleax]

Fwiw, #15448 is a pretty fresh crash coming from async_hooks, although there seems to be more going on in it.