Leader Election with S3 Conditional Writes
morling.dev
morling.dev
I didn’t read past this:
> Systems like Apache ZooKeeper or Postgres (via Advisory Locks) provide the required building blocks for this
Zookeeper is the original sin. Convincing a whole generation of programmers that distributed lock are a feasible solution.
This is my biggest pet peeve in distributed systems.
——
And if you don’t believe me, maybe you’ll trust Kyle K. of Jepsen fame:
> However, perfect failure detectors are impossible in asynchronous networks.
Links to: https://www.cs.utexas.edu/~lorenzo/corsi/cs380d/papers/p225-...
That failures are a possibility does not concern me. That the failures are not fully characterized does.
TCP guarantees order [as long as the connection is active] but it is far from being 'perfectly reliable'.
Example: sender sends, connection drops, sender has no idea whether the receiver received.
In other words, it works until it doesn't. The fact that sometimes it doesn't means it's not perfect.
TCP is a great tool but it doesn't violate the laws of physics. The 2 generals problems is and will always be impossible.
Rounding down 'usually works' to 'it's not perfect and we need to handle edge cases' is how you build dependable systems.
Your first comment seemed very much in the first camp to me.
There’s a relativity-like issue where it’s impossible to have a globally consistent view of time.
See IR2 for how to synchronize physical time in a distributed system.
“(a) If Pg sends a message m at physical time t, then m contains a timestamp Tm= C/(t). (b) Upon receiving a message m at time t', process P/ sets C/(t') equal to maximum (Cj(t' - 0), Tm + /Zm).”
I think the formula is saying that the clocks will only ever increase (i.e. drift upwards). If so, then you could imagine two processes leapfrogging if one of them sends a message that bumps the other’s clock, then that one sends a message back that bumps the first.
But I’m curious how it behaves if one of the clocks is running faster, e.g. a satellite has a physically different sense of time than an observer on the ground.
Also note the paper claims you can’t get rid of clocks if you want to totally order the events.
1ms is an extremely long time in computer land where decisions are made in nanoseconds. There are 1000 nanoseconds per millisecond.
The order of events isn't important here. We only care about the 'last event' which everyone can agree on; they just can't agree on when it happened.
In other words, they can all agree that there is a leader; we just don't know if that leader is still alive or not. The best thing to do is simply make the leader election deterministically random:
1. 'seed' the file with a 'no leader' file.
2. at randomly short intervals (max latency you want to deal with, so like once every few seconds or so), try to claim the file by using the hash of the current file as your etag. The winner is the leader.
3. Once there is a leader and you are the leader, every N seconds, update the file with a new number, using the hash of the current file as the etag.
4. If you are not the leader, every N*(random jitter)*C (where C>N*2, adjust for latency), attempt to take the file using the same method above. If you fail, retry again in N*(random jitter)*C.
5. If the leader is taken, you are not the leader until you've held leadership for at least N seconds.
This doesn't necessarily remove the clock. However, most -- if not all -- modern clocks will agree that the length of a second is within a few nanoseconds of any other clock, regardless of how accurate its total count of seconds is since the epoch. It also probably doesn't work, but it probably isn't that far away from one that does.
For some definition of impossible, given that many systems utilise them effectively. Not all corner cases or theoretical failure modes are relevant to everyone.
For example bicycle wheels are bistable systems, but they usually stay in their useful state so it does not matter in practice.
I too hate this. Not just because the edge cases exist, but also because of the related property: it makes the system very hard to reason about.
Questions that should be simple become complicated. What happens when the distributed locking system is down? What happens when we reboot all the nodes at once? What if they don't come down at exactly the same time and there's leader churn for like 2 minutes? Etc, etc.
Those questions should be fairly simple, but become something where a senior dev is having to trace codepaths and draw on a whiteboard to figure it out. It's not even enough to understand how a single node works in-depth, they have to figure out how this node works but also how this node's state might impact another node's.
All of this is much simpler in leaderless systems (where the leader system is replaced with idempotency or a scheduler or something else).
I very strongly prefer avoiding leader systems; it's a method of last resort when literally nothing else will work. I would much rather scale a SQL database to support the queries for idempotency than deal with a leader system.
I've never seen an idempotent system switch to a leader system, but I've sure seen the reverse a few times.
> I too hate this. Not just because the edge cases exist, but also because of the related property: it makes the system very hard to reason about.
I think this is a huge problem with the way we’re developing software now. Distributed systems are extremely difficult for a lot of reasons, yet it’s often or first choice when developing even small systems!
At $COMPANY we have hundreds of lambdas, DocumentDB (btw, that is hell in case you’re considering it) and other cloud storage and queuing components. On call and bugs basically are quests in finding some corner case race condition/timing problem, read after write assumption etc.
I’m ashamed to say, we have reads wrapped in retry loops everywhere.
The whole thing could have been a Rails app with a fraction of the team size and a massive increase in reliability and easier to reason about/better time delivering features.
You could say we’re doing it wrong, and you’d probably be partly right for sure, but I’ve done consulting for a decade at dozens of other places and it always seems like this.
The older I get, the more I think this is a result of Conway's law and that a lot of this architectural cruft stems from designing systems around communication boundaries rather than things that make technical sense.
Monolithic apps like Rails only happen under a single team or teams that are so tightly coupled people wonder whether they should just merge.
Distributed apps are very loosely coupled, so it's what you would expect to get from two teams that are far apart on the org chart.
Anecdotally, it mirrors what I've seen in practice. Closely related teams trust each other and are willing to make a monolith under an assumption that their partner team won't make it a mess. Distantly related teams play games around ensuring that their portion is loosely coupled enough that it can have its own due dates, reliability, etc.
Queues are the king of distantly coupled systems. A team's part of a queue-based app can be declared "done" before the rest of it is even stood up. "We're dumping stuff into the queue, they just need to consume it" or the inverse "we're consuming, they just need to produce". Both sides of the queue are basically blind to each other. That's not to say that all queues are bad, but I have seen a fair few queues that existed basically just to create an ownership boundary.
I once saw an app that did bidirectional RPC over message queues because one team didn't believe the other could/would do retries, on an app that handled single digit QPS. It still boggles my mind that they thought it was easier to invent a paradigm to match responses to requests than it was to remind the other team to do retries, or write them a library with retries built in, or just participate in bleeping code reviews.
Haha I've seen this anti-pattern too (although I think it's in the enterprise patterns book??). It would bring production to a grinding halt every night. Another engineer and I stayed up all night and replaced it with simple REST API.
Atomic Broadcast (via Paxos or RAFT) does not depend on partial synchrony assumptions to maintain its safety properties.
Your internet or intranet networks are definitely asynchronous and assuming delays are bound is a recipe for building crappy systems that will inevitably fail on you in hard to debug ways.
> leader election will only ever be eventually correct... So you’ll always need to be prepared to detect and fence off work done by a previous leader.
I could not get past the point where you promulgate the idea that ZK can be used to implement locks.
Traditionally a 'lock' guarantees mutual exclusion between threads or processes.
"Distributed locks" are not locks at all. They look the same from API perspective, but they have much weaker properties. They cannot be used to guarantee mutual exclusion.
I think any mention of distributed locks / leader election should come with a giant warning: THESE LOCKS ARE NOT AS STRONG AS THE ONES YOU ARE USED TO. Skipping this warning is doing a disservice to your readers.
[1] https://en.wikipedia.org/wiki/STONITH
[2] https://docs.oracle.com/en/operating-systems/oracle-linux/8/...
Atomic Broadcast guarantees order of delivery. It does not (cannot) guarantee timing of delivery. Which is what people want and expect when using distributed lock / leader election.
If you're saying "what if everything's down and we never get responses to our leadership bids", then yeah, the data center could burn down or we could lose electricity, too.
Process 1 receives: 3:00 [1] "P1 is leader" 3:01 [2] "P2 is leader"
Process 2 receives: 3:00 [1] "P1 is leader" 4:00 [2] "P2 is leader"
This is perfectly valid Atomic Broadcast. Order is maintained.
However from 3:01 to 4:00PM you have 2 leaders (or 2 processes holding the lock).
Don't use ABCast to do locking / leader election for your "user-space" application!
At some point distributed systems that work well vs others that do not is an “art of tuning timeouts and retries”.
Also nothing in production is perfect - so we should consider failures always when writing code in distributed systems and the impacts.
And we will still make mistakes…
'Technically' intractable problems are solvable just fine in a way that is almost as useful as solving them completely if you can achieve one of two things:
* Reliably identify when you've encountered an unsolvable case (usefulness of this approach depends on the exact problem you're solving).
or
* Reduce the probability of unsolvable cases/incorrect solutions to a level low enough to not actually happen in practice.
'Technically' GUIDs are impossible, reliable network communication (TCP) is impossible, O^2 time complexity functions will grow to unusably large running times - but in practice all of these things are used constantly to solve real problems.
I've seen way too many systems where people assume TCP == perfectly reliable and distributed locks == mutual exclusion. Which of course it's not the case.
I know you linked the first paper, but yeah candidly I don’t wish to read the full paper
Don’t mean this sarcastically whatsoever. Genuine interest.
When programmers think of locks, they think of something that can be used to guarantee mutual exclusion.
Distributed locks have edge cases where mutual exclusion is violated.
Implementation does not matter.
e.g. imagine someone shows you a design for a perpetual motion machine. You don't need to know the details to know it doesn't work! It would violate the laws of physics!
Similarly, anyone telling you they created an implementation of a distributed lock that is safe, is claiming their system breaks the laws of information theory.
"Distributed locks" are at best contention-reduction mechanisms. i.e. they can keep multiple processes from piling up and slowing each other down.
[Some] Paxos for example use leader election to streamline the protocol and achieve high throughput. But Paxos safety does NOT depend on it. If there are multiple leaders active (which will inevitably happen), the protocol still guarantees its safety properties.
I might choose DynamoDB over S3 to implement this but both are fine.
That said, if you have a choice between implementing idempotency and using a distributed lock, you should always opt for the former. It’s far less complex and less error prone.
And I thought S3 also supported that. I was always under impression that S3 at some point reached feature parity with GCS.
Are there any other fundamental features they are not equal on?
This is because you're essentially pushing the problem down to S3, which does its own leader election in a way that is waaaay overbuilt for what we're trying to accomplish here.
But... that doesn't mean it isn't cool. :)
Small objects (especially small hot objects) are actually problematic in S3. They cache them in the keymap instead of retrieving from block storage. But many small objects can quickly blow out a keymap shard. The keymap is designed for high-performance lookups. Because of this, it's also more expensive to scale out this class of 'storage' at the keymap layer than to scale out block storage. And you still have to do quorum writes to block storage and then cache eviction at the keymap.
If you're doing this for slow-moving leader election in a small cluster, fine. But if, for example, you started using this for leader-election among end-user-defined workloads that could scale up/down, you might find yourself on the other side of a call from AWS.
I work at AWS but not on the S3 service team. (Opinions are entirely my own.)
I have little doubt that the team already considered this possibility before releasing the feature. The choice to make new functionality available in a service is considered a “one-way door” that receives a tremendous amount of scrutiny. Once a feature is made available, they will do everything they can to support it as long as possible, even at the cost of convenience to themselves. And even de-emphasized services (keep the lights on) are rarely deprecated completely - existing accounts and customers already using them can frequently continue to do so.
By way of example, a little over a decade ago a famous online streaming company used to upload all of their video masters for transcoding. This process involved a huge upload and a huge workload, followed by a significant drop in activity.
The problem was that AWS had to provision for peak usage instead of average usage. This resulted in a situation where the peak-to-average ratio was very high for just one or two customers.
To address this issue, the solution was to incentivize these customers to spread out their workload more evenly over time, at least until they were no longer the largest driver of peak/avg.
This is also why things like Reserved Instances and Spot Instances exist.
(Source: I was on the S3 team. Opinions my own, etc.)
> “I never said they would kill the feature. I said they would shut down the behavior.”
Seems "it" is standing in for a lot of nuance here.
With these three sentences side by side, still took a while to see how you might be referring to the "everyone using" behavior instead of the "S3 purpose" feature! Usually given ambiguity the nearest plausible reference wins.
Since in tech "shut it down" is more often systems than humans and that was the nearest reference "it" could refer to, took some doing to see how your assertion "I said the behavior" could be accurate!
We can continue to debate whether or not I (who worked on that team and generally understands what goes into building scaled cloud services, having built several myself) understand how a cloud provider responds to customers using your service in a way you clearly didn't intend, or we can move on with our day.
As others have pointed out, it's probably not a noticeable cost and, in fact, the fixed costs associated with setting something up yourself would far outweigh what you're paying to use S3 for this purpose.
Part of me just dies inside when I think of all the stuff needlessly happening behind the scenes, given it's not actually being used for storage. I mean, it's called Simple Storage Service.