Comment by andrewcooke | Hacker News Reader

Comment by andrewcooke | Hacker News Reader

Parent Full thread

thanks. given that, can anyone answer the following?

- if there's no intermediate queuing, how are "overloads" handled? what if the system can't handle a transient load? is there some kind of flow control signalling back up the path to control the rate (i assume not!)?

- what happens if a bolt accepts data then crashes before processing. is that data resent? if so, how does the system handle bolts that are not idempotent? (the problem is that if you assume that data is not handled until the handling process terminates then you may send duplicates; if you assume that data are handled if they are received then you lose data on crashes).

- how can "fields grouping" scale? surely if you add more bolts they won't receive any messages unless a new field appears.

- is there some kind of automatic restarting? how are failures handled? [ok, this is handled by supervisor nodes and is described in the link above. although it's difficult to understand how they can be stateless. what happens to the managed processes if a supervisor dies?]

thanks again! [edit: thanks for the great answers; should have guessed the consistent hashing one ;o]

You generally feed a Storm topology with a queueing system at the source. For example, we use Kestrel to feed Storm topologies. There's a setting called "TOPOLOGY_MAX_SPOUT_PENDING" which controls how many tuples can be pending on any given spout task at a time. This is more than sufficient for controlling bursts of data. If things back up, they will queue on the queueing server (s). In the future (not for a few months, at least), Storm will have auto-scaling where it will automatically scale the topology to the data.

If a tuple fails to be processed, the tuple(s) that triggered that tuple are replayed from the spout. See https://github.com/nathanmarz/storm/wiki/Guaranteeing-messag... for more info on that. In that sense, Storm is an at-least-once delivery system (but messages are sent more than once only in failure scenarios). It is up to you to architect your systems to handle this. The approach I take is to build systems using a hybrid of batch processing (Hadoop) and realtime processing (Storm). With batch processing, you can run idempotent functions even with duplication, which lets you correct what's happening at the realtime layer. In that sense, Hadoop and Storm are extremely complementary. Here are some slides from a presentation I gave about this technique: http://www.slideshare.net/nathanmarz/the-secrets-of-building...

Fields grouping uses consistent hashing underneath. So if you redeploy with more parallelism, it scales naturally and easily.

If a supervisor dies, it starts back up like nothing happened. Most notably, nothing happens to the worker processes. All state is kept either in disk or in Zookeeper. All daemons in Storm are fail-fast, and the Supervisor uses kill -9's to kill workers, so that makes things extremely robust.

I hope that answered your questions!

The supervisor/worker thing sounds quite a bit like OTP.

Reply on news.ycombinator.com