> Hell, the author of this exact blog post has picked 10 seconds in all of his examples. That a FUCKING BAD timeout. It's far, FAR too short for many use cases.
It isn't necessarily. It all depends on the use-case. If most of the operations are finishing in 5ms then the probability of something finishing after 10s are rather low - and timing out and retrying early is probably the way to go.
Someone else in this thread recommended setting the timeout to around the P99 time that operations take. I think that's a reasonable starting point, even I might move it towards P99.9.
I worked (and am still working) on adjusting timeouts for systems doing billions of requests/s. One takeaway from that is that the actual value of timeouts is often not too important if you look at one system in isolation. The latency distribution will be rather logarithmic. Most requests might e.g. finish in 20ms. Then you get a P99 at maybe 3 digit ms, and a P99.9 at 10s (example numbers). From there on it will make a minor difference in availability if you now set your actual timeout to 5s or to 120s - it might just be noise along your other error sources.
However it makes sense to align the absolute timeout with timeouts of dependencies. E.g. if you have a chain of
client -> service A -> service B
and service A always times out first, then the client will get an error that service A is broken - but nobody can easily diagnose whether it was service A or service Bs fault. If service B times out first, the client can get an error message that indicates that. Therefore it makes sense if upstream service timeouts are shorter (even if only by a second).
In the same model if the client times out first the services actually do only observe the client dropping the connection. They don't know whether the client timed out or cancelled the operation for other reasons. And therefore they might also not record that something in the service is actually not ideal. For that reason I would recommend setting client timeouts higher than service timeouts (if you are aware of them).
However there is yet another exception to this thing, which are TCP connection timeouts. If you can configure them separately, it makes sense to have those rather low and performing multiple retries. That can improve overall latency, since dropped SYN packets will only be retried by the OS after 1s.
TCP takes 3-4 seconds to detect a dropped packet and retransmit. That puts an absolute minimum baseline of 5 seconds on timeout just to be able to send or receive a packet.
If you're doing anything across datacenters, you have to take a minimum baseline of 10 - 15 seconds to account for extra latency on top.
If you do billions of requests/s I bet you don't care that requests fail? You probably can't even see that requests are failing because you'd have no logging, too expensive at this scale.
I do financial systems, most load typically doesn't go above 1k/s, but every request matters because a dropped request is a dropped payment, possibly tens of millions of dollars lost! There is a ton of issues caused by having too low timeouts set by developers (anything below 30 seconds). I had to reconfigure a ton of systems and libraries to have higher timeouts and ignore configuration passed by developers.
> If you do billions of requests/s I bet you don't care that requests fail? You probably can't even see that requests are failing because you'd have no logging, too expensive at this scale.
That's an assumption. Our customers care a lot. And we have sufficient monitoring in place.
All recommendations I provided above where about maximizing availability, and provided based on the experiencing of improving the experience for lots of users.
> I do financial systems, most load typically doesn't go above 1k/s, but every request matters because a dropped request is a dropped payment, possibly tens of millions of dollars lost!
Without knowing much more about your system: If you are losing that amount of money for failed requests (which can e.g. happen due to random network blips) you are doing something wrong. You should invest into different strategies than increasing timeouts.
It's inherent to payment systems to "lose" money on a failed request. A transaction involves two sides, either both sides agree that the transaction is completed or money is lost/duplicated.
If the client decides to drop (timeout) and consider the transaction cancelled, while the server is processing it and will consider it done. That's a catastrophic issue that needs to be addressed. It is one of the most common bugs I've seen in the wild (root cause: too short timeouts).
How to make highly critical systems reliable enough in the face of hardware and software issues is a complex topic. At this level this involves a holistic approach to get every component to cooperate together (timeout is a minor example). A HUGE amount of work is to detect errors, and more importantly to propagate errors across diverse stacks (software should be aware of database errors, services should detect other services failing).
This doesn't make sense to me. If there is a real risk that a timeout can happen (and there always is) then the payment system should be implementing a two phase commit.
I don't know what the Byzantine generals problem is.
Two stage commit is important because it has:
1) Predefined transaction id prior to final submission that allows you to validate the status, so if your request to commit gets 503'ed or you get a timeout you can reliably query to know if it was processed or not
2) Unlimited resubmissions of the final commit. It doesn't matter if I perform the final commit api request 1 time or 100 times, it will never cause a duplicated transaction to occur. So if I get a timeout or a 503 I can resubmit knowing that if my original commit request went in my new submit will be a no-op, and if my last commit request didn't get processed then this time it hopefully will be processed.
This pattern isn't just a payments pattern thing either. This is heavily used in distributed systems where failures can occur. UPS' API used to use this as well so you could be sure that you don't pay for duplicate shipping labels or cause duplicate shippments.
The Byzantine general problem is the field of research dealing with consensus/consistency issues like what we discuss here. The baseline is that there are two generals on a battlefield trying to coordinate an attack, they send messengers to communicate but any message might be lost or intercepted. The problem is proven to be unsolvable so let's not go heads in assuming you can be sure of any outcome ;)
Taking longer does not solve the byzantine generals problem. The difference here is that the role are asymmetrical: once the bank receive your order for a transaction it does not need to check that you know whether the order was correctly received; the bank can simply perform the transaction and them best-effort let you know of what happened.
Isn't it better to make it idempotent? The risk is that the client might accidentally make the same transaction twice if the first attempt looks like it failed.
Make the client include the id of it's last known transaction and only apply the transaction if it's up to date, otherwise tell the client to refresh and try again.
The second stage is idempotent (which is why it works), but the purpose of the first stage is to make sure both sides have an agreed upon idea of the uniqueness of the transaction that's about to take place.
For instance, if I want to generate a shipping label that goes from my house to your house and I do two attempts, how does the receiving service know if I made two distinct attempts (I want to ship 2 similarly sized items) or if a transient error occurred in between making me attempt a re submission?
You solve this by creating an inactive request with the criteria (shipping label from my house to your house). This step is not idempotent but that's OK, because if I resubmit I just create a 2nd inactive request that may never actually be finished.
The second step is to say "this request is good and I want to proceed with it". That step is idempotent and marks the existing request as not just inactive but puts it in an active state.
A shopping cart flow is a user managed 2 stage commit (review your cart, submit the cart order). No matter how many times I submit my order it won't cause duplicate orders because I'm submitting a specific shopping cart.
UPS, Paypal, and others just use a computer/api-managed 2 stage commit
You can't always rely on a client generated ID, because you would have to know that the client id is unique enough. The server is the only one who can really generate a transaction id that it knows is globally unique and efficiently queryable in its backend.
It's not mutually exclusive. You can do two stage commits with the second stage being idempotent.
The practical risk is that this puts a ton of complexity on the client, to keep track of states and perform some follow up actions. The added complexity means more bugs and each additional step can fail hence compounding the problem rather than solving it.
This doesn’t address much of your comment, but I work on systems with millions of req/s, but even with billions you can still sample to do logging and monitoring. But you’re right, we don’t care that every request works, just that 99.9% do.
> recommended setting the timeout to around the P99 time that operations take. I think that's a reasonable starting point, even I might move it towards P99.9.
Why would you intentionally drop 1% or even 0.1% of requests?
What's the process of finding P99? Is it taking a bunch of samples, getting the standard deviation, and then calculating the value at which 99% of all possible samples would fall?
This also assumes O(1) I assume...
(I'm thinking of how to apply reasonable timeouts to background celery tasks)
It isn't necessarily. It all depends on the use-case. If most of the operations are finishing in 5ms then the probability of something finishing after 10s are rather low - and timing out and retrying early is probably the way to go.
Someone else in this thread recommended setting the timeout to around the P99 time that operations take. I think that's a reasonable starting point, even I might move it towards P99.9.
I worked (and am still working) on adjusting timeouts for systems doing billions of requests/s. One takeaway from that is that the actual value of timeouts is often not too important if you look at one system in isolation. The latency distribution will be rather logarithmic. Most requests might e.g. finish in 20ms. Then you get a P99 at maybe 3 digit ms, and a P99.9 at 10s (example numbers). From there on it will make a minor difference in availability if you now set your actual timeout to 5s or to 120s - it might just be noise along your other error sources.
However it makes sense to align the absolute timeout with timeouts of dependencies. E.g. if you have a chain of
client -> service A -> service B
and service A always times out first, then the client will get an error that service A is broken - but nobody can easily diagnose whether it was service A or service Bs fault. If service B times out first, the client can get an error message that indicates that. Therefore it makes sense if upstream service timeouts are shorter (even if only by a second).
In the same model if the client times out first the services actually do only observe the client dropping the connection. They don't know whether the client timed out or cancelled the operation for other reasons. And therefore they might also not record that something in the service is actually not ideal. For that reason I would recommend setting client timeouts higher than service timeouts (if you are aware of them).
However there is yet another exception to this thing, which are TCP connection timeouts. If you can configure them separately, it makes sense to have those rather low and performing multiple retries. That can improve overall latency, since dropped SYN packets will only be retried by the OS after 1s.