Job queues are deceptively tricky
typesanitizer.com127 points by ingve 3 days ago
127 points by ingve 3 days ago
The UNIX pipe really is an incredible concurrency invention which is not well understood, and attempts to work around its features turn into bugs.
A buffer to accumulate data that blocks when it’s full allows you to handle bursty loads. It solves the back pressure problem of readers and writers operating at different speeds. It doesn’t over consume resources.
It also solves the architecture problem of when to trigger work. Both consumer and producer act on the pipe imperatively, rather than one end being imperative and the other being a declarative graph of callbacks (all those “reactive” libraries).
Even using a term like “back pressure” is a tell to me that someone is confused snd doing something architecturally wrong.
The UNIX pipe has the (for many systems) undesirable negative property of losing data in the pipe when the receiving process terminates: Anything in the kernel buffer of the pipe gets lost.
Since the pipe is generally unidirectional, the sending process has no way of knowing whether the receiving process has received, or even more successfully processed, anything sent. For that, one needs to make a pipe in the opposite direction and that is not as easy anymore; it also requires building your own protocol to identify and acknowledge sent work items.
> the pipe is generally unidirectional
So are most message queues.
If the producer needs to know when the consumer has finished work, that's not a queue; that's an RPC.
And to have a hope of solving that problem you need to restore producer and consumer thread states without encountering that problem.
Bad take. Most of the different behaviors you criticize about message queue systems arise from them needing to a) distribute work b) durably, c) over the network.
Most message queues want to be able to distribute work to multiple parallel consumers, either for performance, redundancy, re-deployment of the consumers, and so on. Like, sure, you can do that with named pipes, but they're not durable; their state can be lost in the event of a crash or reboot.
Other than like ... in-process, thread-to-thread message queues, almost all message queues are used because they provide some form of durability (either by persisting messages to disk, or just by virtue of running on a separate server/process than much more crash-prone producers/consumers).
> imperatively, rather than one end being imperative and the other being a declarative graph of callbacks (all those “reactive” libraries).
MQ client libraries are heavily reactive because they need to work with MQs that are network services. If every consumer that ever wanted to fetch a message had to issue a network RPC poll/timeout cycle, and they were all doing that constantly, that'd be both a lot of traffic for the MQ to handle and a lot of network chatter. Some MQs do use (or support) the RPC poll model, but especially among the more performant ones, most of them are push-based so clients can just sit in an epoll/select statement waiting for the message queue to send them traffic.
That said, you're totally right that plenty of MQ client frameworks/libraries are way over complicated and want to own your whole program's entry point and design. Callbacks are probably a necessary outgrowth of the underlying network behavior, but tortured whole-program-event-loop-ownership architecture is not.
> Even using a term like “back pressure” is a tell to me that someone is confused snd doing something architecturally wrong.
I don't think your other points support this; you said yourself that pipes' bounded nature is good, so what's the problem here?
I just talked about pipes as a useful conceptual tool. You’re assuming what applications I think they apply to. But I’ll do my best.
One of my takeaways is that if you do not handle on data transfer in the http request/response then you are choosing to exit the pipe model. Once you do that, yes all kinds of crazy async problems exist which require very complex tools to wrangle.
See the sibiling comment from Google engineer. It’s better to do work synchronously than delay it until later.
So now we get your list of requirements that MQ is for. Ok but how did we get here? What problem are we trying to solve? The answer is usually a performance problem for which this is a bandaid.
Another possible model for this problem is email. You can find some great threads here about large amounts of infrastructure being replaced with Unix email tools. Although I have yet to try it!
> If every consumer that ever wanted to fetch a message had to issue a network RPC poll/timeout cycle, and they were all doing that constantly, that'd be both a lot of traffic for the MQ to handle and a lot of network chatter
This makes no sense to me. A while loop blocked on a socket read with a producer with a write does not generate any extra network “chatter”.
Anytime you are using callbacks you are choosing not to have a thread with imperative logic.
> I don't think your other points support this
Correct that was not a conclusion.
What I’m trying to say is if you want to send data between two systems and you have producers and consumers at different rates, you don’t have to do anything special. The UNIX kernel is designed to solve this problem.
So when I hear that, my inclination is that we don’t understand the capabilities UNIX already offers and I’ve yet to be wrong.
> The answer is usually a performance problem for which this is a bandaid.
I agree that if that's the problem you're trying to solve, MQs are a poor stop-gap and not a general-purpose solution.
But many (most?) MQs aren't deployed for that reason; rather, they're used to either a) defer work whose latency characteristics are incompatible with the producer's runtime (e.g. sending email which might take minutes/hours to be accepted upstream can't be done synchronously from an HTTP request handler), b) handle work which doesn't need to be done transactionally and which has a high probability of needing to connect to resources whose uptime you don't control (queues make retries easy), or c) work that needs to be batched or otherwise completed on a schedule or dimensionality that's not available in the producer.
> A while loop blocked on a socket read with a producer with a write does not generate any extra network “chatter”.
True, and some MQs support that pattern, but it's rare for the same reason that most network server applications' core loop doesn't wrap a blocking read. While-blocking-read plays poorly with timeouts, restarts, and multiplexing/wait-any. This isn't unique to MQs.
> Anytime you are using callbacks you are choosing not to have a thread with imperative logic.
I mean, it's pretty easy in most languages/runtimes to turn a callback into an imperative wait. When it comes to network servers specifically, I think exposing the lowest-runtime-overhead model as the primary API (selectors and callbacks) and letting users add control flow primitives on top of that is preferable, but reasonable minds can differ here. This also isn't unique to MQs.
> if you want to send data between two systems and you have producers and consumers at different rates, you don’t have to do anything special. The UNIX kernel is designed to solve this problem.
For ephemeral/can-be-lost local traffic, you're partly right (pipes/fifos are a simple and widely available tool that can help with this, though you have to layer some protocol-ish logic on top to make them work reliably between consumers that expect particular structures and don't speak ASCII/newlines alone). For durable local traffic, or remote traffic, UNIX doesn't have a prebuilt primitive. TCP tx/rx window sizes aren't surfaced such that buffers and backpressure can be reliably interpreted by userland to make delivery decisions.
Thanks for high quality reply.
> latency characteristics are incompatible with the producer's runtime
> handle work which doesn't need to be done transactionally and which has a high probability of needing to connect to resources whose uptime you don't control (queues make retries easy
I agree. I think you do need to break the request/response pipe for this use case.
I traditionally have treated this as a state in a database. I know there are limitations and I can see MQ being a solution. I want to try the email idea sometime.
I do think queues are a poor technology for dealing with performance problems in a web stack.
> This also isn't unique to MQs.
Agreed. My ranting about async and callbacks is an ongoing project.
I think pipes solve producer/consumer problems in a way most engineers don’t appreciate.
> While-blocking-read plays poorly with timeouts, restarts, and multiplexing/wait-any.
The primary problem solved by not using blocking primitives is to try to free up OS resources from threads (green threading). Why would an internal queue have that problem? It’s not accepting arbitrary connections.
I think the the second part was kinda obvious? The moment I read this:
> If you’re anything like me, you would probably have said Parallel Spawn, Prefer New, and Wait are perhaps defensible, whereas Prefer Old feels weird/backward.
it was pretty obvious I was not anything like him. My intuitive answers are pretty different.
- Parallel Spawn is useful, but it's orthogonal to the rest. Even if you have 4 workers, you'll still have to worry about hitting concurrency limit once you have enough jobs. What is it even doing in this list?
- Wait is very useful for non-scheduled tasks: if user uploaded 100 files to process, you better process them all. Sometimes you need limit, sometimes you do not (let them queue for a while until devops notices and either allocate more workers or clear them and has some harsh words with consumer).
For scheduled tasks, "Wait" seems much less useful. I can come up with a reason but they are all somewhat convoluted - perhaps you are hitting 3rd-party service, and it has a quota, so you've decided to use scheduler to avoid hitting ratelimits?
- "Prefer Old" is normally the best way. You repack takes 3-8 hours, so you set your timer to "every 1 hours, skip if running already" and you can be sure that your job finishes and the next one will start.
- "Prefer New" seems almost useless. You've already spent all this effort doing the job, why are you cancelling it and throwing away the results? If you want to add a timeout, add a timeout, preferably to specific operation. For example, if there job starts by fetching data, and this fetch can be super slow, use 'Prefer Old' and put a timeout on the fetch part. This way your job won't be interrupted if the fetch just finally succeed minutes before next scheduled interval hit.
Oh, and re "If the Prefer Old semantics are not offered, you can’t really emulate them using the two primitives of regular scheduling and limiting concurrency." - you totally can. Set concurrency to 2, and add as a first thing: "fetch the list of jobs running; if there is anyone except me, exit right away".
Really, not that tricky at all.
At Google we actually had 'prefer new' (ie a stack instead of a queue) for certain jobs, that were likely no longer useful after some time had passed; and where less and less useful the longer you waited until you started.
One example was running certain ad auctions when rendering websites, or something like that. You don't want to delay serving the side, if the ads are delayed.
So you have a certain wall clock time budget until the rest of the page is assembled to be sent to the user, and if you can fit your ad-serving in there, that's good.
If you have more work than you can currently handle, then it makes sense to continue with the newest open request after you handled the previous request.
The actual system was a lot more complicated, and combined a short, bounded queue on the inside with a large stack on the outside or something like that.
Similar considerations can apply, when you are one of many competing market makers for some financial assets on an exchange.
Basically, if you have a situation where serving quick is a lot more important than the distinction between late and very late (or even dropping the request).
That's just another word for priority queueing.
A priority queue where the last thing to enter the queue is the highest priority is a stack, yes.
> - "Prefer New" seems almost useless. You've already spent all this effort doing the job, why are you cancelling it and throwing away the results?
The obvious use-case is if the result depends on a state and the state has changed since the job has started. Examples would be updating the landing page when a new article has come in (you don't need the outdated landing page w/o the new article) or if you did a code change while compiling (assuming it's not a commit).
Major lesson from when I worked on Google Search indexing is that queues have a lot of hidden complexity and can make your outages much longer than they need to be. We had a big project to get rid of a bunch of queues by just scaling up our synchronous backends and making them faster.
Care to share more about the issues?
Not OP but also worked on Google Search once upon a time. I'm not sure if I'm remembering the same issues as OP, but basically the two biggest issues are:
1.) What they do to your 95th percentile latency. Users are often very sensitive to tail latency: a service that responds in 150ms 19 times and then takes 2s on the 20th is still perceived as annoyingly slow. With job queues, the reason for that slowness could be as simple as "it was the 20th request to arrive during a period of high demand, and backends couldn't keep up". The whole point of queues is so you can gracefully handle this case without overprovisioning your backends by a factor of 20x, but if the user is still going to consider this a miss anyway, you have to overprovision the backends anyway. There isn't really another way to handle this other than having spare backend capacity. Also note that in many cases the user hitting "refresh" doesn't cancel the existing queued job, it just adds another one to the queue. Which brings us to...
2.) They can turn simple failures into cascading failures. There were several postmortems that went something like "Service X became overloaded because of an unexpected flood of requests, leading to several individual replicas shutting down. This led to more requests being routed to the remaining replicas, which overloaded them too and led to all of Service X going down. When SRE attempted restart Service X, requests queued in the job queue were all retried en masse, which led to an overload of the partially-restarted service and a subsequent failure. SRE had to limit requests upstream and manually drain all job queues and bring Service X back cluster by cluster to restore service health."
The root principle here is that any distributed system needs a concept of backpressure. When critical downstream dependencies are overloaded, they need to pass this information back up the stack to the entry point, which needs to start denying requests from the user or do a simpler fallback that doesn't put load on the overloaded service. Naive queueing does not work, because the requests are still sitting there in the queue waiting to overload the downstream service once it becomes available again.
You can bolt backpressure onto a job queue system (by eg. rate-limiting requests to a service that has just come back up, or rate-limiting based on response time, and/or falling back to simpler algorithms), but at that point, it's a backpressure system, not a job queue. The semantics are very different from a system that guarantees eventual delivery, just not sure when. You need to be able to handle partial failures and adapt with different algorithms at multiple points within the system.
I also worked on services at Google, both as an SRE and a SWE.
I think one subtlety worth bearing in mind here is that Google, at least during my tenure (2006-2014) didn't actually have a proper message queueing system outside of Gmail, which was used only for email delivery. It offered engineers:
1. RPC with infinite backoff/retry (a fun default).
2. Batch jobs that processed files.
but there was no equivalent to a standard enterprise MQ product like ActiveMQ, Artemis, Oracle AQ and so on.
So when we talk about "job queues" it's worth being very precise about what is meant. A standard enterprise message queue broker doesn't have problems with overload because workers pop work off the queue at whatever speed they can operate. The problem of cascading failures and services coming back only to be immediately overloaded again was a problem caused by the design of Google's infrastructure, in which RPCs would back up in memory in the clients and be retried in a loop until the service came back. So of course this required a lot of custom work to create proper backpressuring, which wasn't normally done and especially not on interactive serving paths, leading to this kind of repetitive failure mode. It was all made harder by the practice of making everything fully async, so thread pool sizes also didn't exert backpressure.
Looking back on my time there, one of the pieces of "normal" enterprise infrastructure that I really think Google could have benefited from was a proper JMS compliant scalable message broker. You can buy these - the Oracle Database has one - but Google neither bought one nor developed its own.
Probably they have long since rectified this oversight.
Pre-Emptive scaling ala erlang can help with scenario one somewhat, if the jobs aren’t locked on some resource. For example, on my erlang system 20 would all run just each slightly slower as they get a smaller amount of scheduler reductions each.
It’s a hard/interesting problem, and harder still once you’re running across a lot of machines — but if they all get slower under the load it turns out that it’s easier to scale / work out a good balance of idle capacity to guarantee x time sla under x requests.
Fast ramp up for additional capacity is important too, but less so if you know to start the process once median execution time drops to some % of your worst target
> Pre-Emptive scaling ala erlang can help with scenario one somewhat, if the jobs aren’t locked on some resource. For example, on my erlang system 20 would all run just each slightly slower as they get a smaller amount of scheduler reductions each.
Well, memory is one of these resources that you are often locked on.
If you have enough memory, running 20x the load just goes 20x slower. But if memory is congested, then this can go arbitrarily slower than running your jobs one after another. Eg when you are swapping to a spinning disk.
I’ve never enabled swap on a production system in ~25 years, so i have to say it’s not really my experience.
Shared memory can be a problem, sure, which is why I don’t generally use that either. Erlang, of course, generally does not use shared memory outside of some cases with ETS etc which must be used carefully but i’d rather solve those problems myself.
Concurrency systems in other languages i’ve written, for example Go, there are ways to architect to avoid it also. I’d rather go slightly slower and copy be value than have to solve mutex contention and trying to make everything atomic and so on. YMMV, I don’t work in HPC just large complex busy systems.
To make my point more abstractly:
Multiple tasks can share the CPU and just get a bit slower. But if you are out of RAM, you are better off running one thing after another.
Whether you hit swap or OOM was a distraction.
Properly scaling queue consumers is a problem I've spent a lot of time on in the last few years. Working on a messaging platform with highly variable traffic, including close to zero during the night, means that capacity provisioning according to the max will be very costly, and lead to a lot of frustration when you are saturated anyway.
Indeed you need backpressure but the traditional methods (CPU usage or similar metrics) are difficult because many consumers aren't high on those metrics --imagine a messaging plaform, pure IO. Also you'd have to tailor to the consumer itself and that's difficult, which is what you mention on the next-to-last paragraph.
In the end I helped solving it by scaling based on queue size and input/output rates, agnostic to the consumer itself, but with the hypothesis that you can scale consumers linearly (or at least monotonically, some sublinearity is allowed). The queue scaler watches for incoming and outgoing traffic on the queue, plus items on the queue itself, and it can scale from 0 to 11 in seconds for gusts, then shutting everything down.
It's a satisfying problem to work on, but its proper solution demanded quite the investigation. Now every queue we've got in the system is managed by this autoscaler -- except when we can't ensure linearity.
> capacity provisioning according to the max will be very costly
I’m skeptical. You can support a pretty massive messaging system with one box.
What did you try? What went wrong?
Similar to what gp mentioned, max latency. You can tune your consumers to burn through the queue and push messages, but when you have latency constraints e.g. a very tight deadline the first few messages in the queue might achieve it but not the 20000th.
When you want to switch from no processing to say 2M messages than need to be sent within 5 minutes it's difficult to appropriately provision without fast scale out.
> Users are often very sensitive to tail latency: a service that responds in 150ms 19 times and then takes 2s on the 20th is still perceived as annoyingly slow.
This reminds me of some research I read about in the 1980s or 1990s on perceptions of the speed of command line commands. If command time varied over a range from nearly instantaneous to say 100 ms fairly uniformly through that range, people would perceive the system as overall being faster when the researchers added a variable delay to all the commands that made them all take 100 ms.
Humans apparently really like consistency.
> When SRE attempted restart Service X, requests queued in the job queue were all retried en masse, which led to an overload of the partially-restarted service and a subsequent failure
...and this reminds me of something else, from around 1983. I was working at a small Unix workstation maker. The guy in the office across the hall found one morning that the battery for the clock on his workstation had died, and the system time had come up after boot as the Unix epoch.
He shut down, put in a new battery, booted, and then set the clock to the current time, 13 or 14 years after the epoch.
Almost immediately his hard disk light came steadily along, and he could hear the disk furiously seeking, and the system became completely unresponsive.
It turned out AT&T cron in the early '80s wasn't smart about time changes. It had tried to all at once every cron job that should have run in the 13 or 14 years that the time just jumped.
The solution to the complexity of any queuing system is to add another queue
The most popular resource manager for job submission and queueing system is Slurm. It's being used in majority of TOP500 supercomputers, and overwhelming majority of the world HPCs [1].
SchedMD the leading developer of Slurm has recently being acquired by Nvidia, while Slurm remaining free and open source, but somehow it's Wikipedia entry is not yet updated accordingly.
[1] Slurm Workload Manager:
https://en.wikipedia.org/wiki/Slurm_Workload_Manager
[2] Nvidia Acquires SchedMD (7 comments):
You solve this simply with two cron jobs, one for weekend and one weekdays.
There is 4th option that looks like a combination of how all three are described here: If J1 is running and J2 gets queued, then when J3 gets queued you cancel J2 but not J1. Prefer the oldest running (well, any running, just don't kill them) and newest queued, which waits on the running one. This is how we had long-running test suites configured on Jenkins/Hudson/buildbot ages ago (though it wasn't exactly a queue, more just a flag that the job needed to be run and the job pulled in the latest state).
Tangential, but when dealing with queues, the first thing you want to do is have a basic grounding of queuing theory, and know whether you're optimising for throughput or worker utilisation (i.e. what are your SLAs and efficiency targets?). IME each goal involves fairly different metrics and scaling rules, so you'll want to know what you're prioritising.
Anyone know a good into to queuing theory?
There are lots of good books and some great vids on youtube, but I'd start with this statement and work backward, because this is the non-obvious thing most bootcamp trained, promoted to CTO don't know:
The single most important lesson from queuing theory for software systems is the non-linear relationship between utilisation and latency.
As system utilisation approaches 1.0 (100% capacity), the average waiting time does not scale linearly, it scales hyperbolically. A system running at 95% utilisation is vastly more fragile and slow than one running at 80%, even though the load difference is minor.
To explain this: if the system is 95% utilized, and a new request comes in, there's a 95% chance the system is already busy and the request has to wait. But after the currently in-progress request finishes, there's still a 95% chance the system is busy (with another request that was already queued behind it) and request X has to wait. After that one finishes, same thing. On average, request X has to wait for about 20 other requests at 95% utilisation - or 5 requests at 80% utilisation - or 10000 requests at 99.99% utilisation. And that's just the mean, not percentiles.
Your math doesn't make any sense.
Which part? 95% utilisation means at any random point in time there's a 95% chance the system is utilized and a 5% chance it is not utilized at that moment, so on average you have to try 20 times to find an unutilised moment?
I remember learning about CSV parsing and how it's conceptually simple, yet beyond the simple , and quotes: the corner cases bloat your parser 10-15x.
Or, you could be a government agency that implemented a method for citizens to upload compliance data - so they have to upload broken CSVs into a broken queue system. Get the worst of both worlds, and deployed as a green-field project in 2021.
It’s pretty simple to make one that’s RFC compliant. The rules aren’t much more than what you said.
Are you talking about trying to interpret malformed data?
I find it very annoying when queue problems break into queue-of-queue patterns like in the `wait`scenario
I haven't modeled it, but I wonder how far you'd get on randomizing the policy choice for concurrency limit 1. Maybe weighted by past results, but bounded to allow it to shift instead of falling permanently into a basin.
Seems like OP has conflated the job queue and the job scheduler here.
The schedule interval vs. hard timeout distinction is the useful bit here. Treating them as one setting makes backlog behavior much harder to reason about.
Nice post. Thanks for sharing.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]