I'm sorry, it took *how* many servers to post a single long message from Ghost to 5k fediverse accounts and handle some replies?
-
@_elena It's a bit weird, yeah! I have nothing but sympathy for their implementation struggles, and I appreciate them being so transparent, it's just that the business implications were a bit startling.
-
Tbh I appreciate the transparency, even when the implications are startling, in part because it draws out other people's similar struggles. It seems like a whole lot of factors conspiring to be a problem—but also some great people working on those factors.
-
Melroy van den Bergreplied to Erin Kissane on last edited by
@kissane @fediversereport @thisismissem well I'm a contributor to Mbin. So try to rewrite Ghost to PHP
. Problem solved. -
Emelia 👸🏻replied to Melroy van den Berg on last edited by
@melroy @kissane @fediversereport mbin uses a queue iirc, which Ghost isn't currently. That's why it's not as performant as it should be
-
I'm not sure that's actually a useful thing to analyze for these sorts of questions? Like I'd be interested in it regardless, but the reason you generally have a queue is not performance.
(Unless by "benchmark" it means "throughput benchmarking" in which case it is useful but it is really just benchmarking the queue's performance and is highly sensitive to it)
-
But even if it lowered the throughput, it's still not done generally for that reason.
You have queues to smooth out traffic spikes, to provide backpressure, and to give resiliency if a process goes down. Not really for anything you'd benchmark?
-
@hrefna @jenniferplusplus @kissane @fediversereport @hongminhee right, but if you're processing synchronously vs asynchronously, then that'll cause a change it throughput and performance.
-
It does, and maybe I'm misunderstanding, I'm just not convinced it matters when the reason you are doing something is not performance related in the first place.
Like NIO is classically slower than synchronous IO in java, but benefits of NIO aren't about performance but about program design and flexibility with different arch (and if the performance difference is enough to matter you probably shouldn't be using java).
-
Like we'd always benchmark apps with the queue, but comparing them to the non-queue scenario seems very weird to me
Because I can't think of a reason you wouldn't already want a queue in a distributed system like this for non-performance reasons
So of course you benchmark, but only the case with a queue means anything and then you tune your queue against different scenarios, or compare types of queues. If that makes sense
-
@thisismissem @hrefna @jenniferplusplus @kissane @fediversereport @hongminhee it shouldn't really unless the queue itself becomes a bottleneck or conversely enables better fanout
but usually it just shouldn't
-
@thisismissem @hrefna @jenniferplusplus @kissane @fediversereport @hongminhee note that the queue mainly impacts perf when it is the bottleneck or when it enables a better architecture
-
@hrefna @jenniferplusplus @kissane @fediversereport @hongminhee
I think the benchmark is more to show how much faster the activitypub processing is when you defer that work by using a queue vs not, i.e., the HTTP requests duration drops significantly most likely; but autoscaling on open http requests / connections probably isn't right for node.js, instead you'd want a combination of event loop lag, mem/cpu, and open requests
-
@noah @hrefna @jenniferplusplus @kissane @fediversereport @hongminhee
So specifically here it's the HTTP throughput to measure, not the "I processed the activity" throughput, does that make sense? So synchronous work versus add to queue
-
@thisismissem @hrefna @jenniferplusplus @kissane @fediversereport @hongminhee ah, so you want to measure the impact of moving this out of the critical path
that makes a lot more sense
-
Jenniferplusplusreplied to Noah Kennedy on last edited by
@noah @thisismissem @hrefna @kissane @fediversereport @hongminhee
I think Emilia means that AP message delivery is currently happening in-band with the request-response cycle. That manifests as slow responses, and drives autoscaling to continue serving demand.The queue (I assume) moves delivery out of band, so the response can complete earlier.
-
Noah Kennedyreplied to Jenniferplusplus on last edited by
@jenniferplusplus @thisismissem @hrefna @kissane @fediversereport @hongminhee i understand now, i replied elsewhere in the reply tree
testing the impact of moving this out of the critical path makes a lot more sense
-
@jenniferplusplus @noah @hrefna @kissane @fediversereport @hongminhee yup, exactly.
-
@jenniferplusplus @thisismissem @hrefna @kissane @fediversereport @hongminhee as soon as i saw her clarification i understood what she was doing, i've done the same thing before a few times
-
Yeah, I get that, I think my confusion is it feels to me a bit like saying:
"If you are going across the water, you can cross Lake Pontchartrain faster in a motorboat than a cybertruck"
You absolutely can. But even if you couldn't, does it matter?
-
@thisismissem @hrefna @jenniferplusplus @kissane @fediversereport @hongminhee I appreciate this more in depth discussion. I just reread the post and the comments. They don't mention what they are autoscaling on. I agree that node services are a different beast than systems that default to blocking I/O. Knowing that they didn't have queueing to create throttling and back pressure, I have other suspicions about what went wrong.
