Fix Version/s: None
Cassandra has the unfortunate behavior that when things are "slow" (nodes overloaded, etc) there is a tendency for cascading failure if the system is overall under high load. This is generally true of most systems, but one way in which it is worse than desired is the way we queue up things between stages and outgoing requests.
First off, I use the following premises:
- The node is not running Azul
- The total cost of ownership (in terms of allocation+collection) of an object that dies in old-gen is much higher than that of an object that dies in young gen.
- When CMS fails (concurrent mode failure or promotion failure), the resulting full GC is serial and does not use all cores, and is a stop-the-world pause.
Here is how this very effectively leads to cascading failure of the "fallen and can't get up" kind:
- Some node has a problem and is slow, even if just for a little while.
- Other nodes, especially neighbors in the replica set, start queueing up outgoing requests to the node for rpc_timeout milliseconds.
- You have a high (let's say write) throughput of 50 thousand or so requests per second per node.
- Because you want writes to be highly available and you are okay with high latency, you have an rpc_timeout of 60 seconds.
- The total amount of memory used for 60 * 50 000 requests is freaking high.
- The young gen GC pauses happen much more frequently than every 60 seconds.
- The result is that when a node goes down, other nodes in the replica set start massively increasing their promotion rate into old gen. A cluster whose nodes are normally completely fine, with slow nice promotion into old-gen, will now exhibit vastly different behavior than normal: While the total allocation rate doesn't change (or not very much, perhaps a little if clients are doing re-tries), the promotion rate into old-gen increases massively.
- This increases the total cost of ownership, and thus demand for CPU resources.
- You will very easily see CMS' sweeping phase not stand a chance to sweep up fast enough to keep up with the incoming request rate, even with a hugely inflated heap (CMS sweeping is not parallel, even though marking is).
- This leads to promotion failure/conc mode failure, and you fall into full GC.
- But now, your full GC is effectively stealing CPU resources since you are forcing all cores but one to be completely idle on your system.
- Once you go out of GC, you now have a huge backlog of work to do that you get bombarded with from other nodes that thought it was a good idea to retain 30 seconds worth of messages in their heap. So you're now being instantly shot down again by your neighbors, falling into the next full GC cycle even easier than originally.
- Meanwhile, the fact that you are in full gc, is causing your neighbors to enter the same predicament.
The "solution" to this in production is to rapidly restart all nodes in the replica set. Doing a live-change of RPC timeouts to something very very low might also do the trick.
This is a specific instance of the overall problem that we should IMO not be queueing up huge amounts of data in memory. Just recently I saw a node with 10 million requests pending.
We need to:
- Have support for more aggressively dropping requests instead of queueing them when sending to other nodes.
- More aggressively drop requests internally; there is very little use to queueing up hundreds of thousands of requests pending for MutationStage or ReadStage, etc. Especially not ReadStage where any response is irrelevant once timeout has been reached.
A complication here is that we cannot just drop requests so quickly that we never promote into old-gen. If we were to drop requests that quickly when outgoing, we would be dropping requests every time another node goes into young gc. And if we retain requests long enough for other node's young gc, it also means we retain them long enough for promotion into old-gen with us (not strictly true with survivor spaces, but we can't assume to target the distinction there with any accuracy).
A possible alternative is to ask users to be better about using short timeouts, but that probably ups the priority on controlling timeouts on a per-request basis rather than as coarse-grained server-side settings. Even with shorter timeouts though, we still need to be careful about dropping requests in places it makes sense to avoid accumulating more than a timeout's worth of data.