[jira] [Commented] (CASSANDRA-3927) demystify failure detector, consider partial failure handling, latency optimizations

["Peter Schuller (Commented) (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-3927?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13210148#comment-13210148 ] 

Peter Schuller commented on CASSANDRA-3927:
-------------------------------------------

Ways I can remember/think of now that I think are long the lines of something to do in the request path (some of it covered by tickets, some of it not; might hunt them down later):

* Consider number of outstanding requests to a node (least outstanding, or a soft variation thereof).
* Make the output of the dynamic snitch probabilistic instead of discrete; a node should not instantly and suddenly flap over in a replica set, but rather a slow host should be appropriately de-prioritized.
** Avoids sudden flapping behavior.
** Avoids e.g. having certain nodes be completely cold for a replica set due to snitching, followed by suddenly taking full traffic and causing poor latencies/timeouts as a result of no cache locality (happens easily with no or very low read repair).
* Data reads to all nodes (or at least multiple, maybe tunable).
* Even when normally not doing data reads to all for efficiency, do so speculatively (as Stu has mentioned in one of the other tickets).
* Be more restrictive about queueing tasks and requests; bound GC/memory footprint impact.
** For example, with RF=3 and a QUORUM request, if one replica is completely backed up - let's not even send a request to that node, but send reads to the other two. This prioritizes requests that really *need* a response from a node.
** For example, put caps on outstanding data beyond that implied by rpc timeout and request throughput.
** For example, spilling over to non-heap might be doable. Think of an off-heap (maybe persistent on disk/SSD) queue.
** Be more aggressive about dropping incoming requests.
** Never ever ever rack up hundreds of thousands or millions of tasks internally in front of a stage.
* Immediately react to the fact that a TCP connection (our only communication mechanism) got dropped (with appropriate adjustments to try to keep TCP connections up regardless of message flow).
* Support "give me data if smaller than X, only digest if not".

                
> demystify failure detector, consider partial failure handling, latency optimizations
> ------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-3927
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-3927
>             Project: Cassandra
>          Issue Type: Wish
>            Reporter: Peter Schuller
>            Assignee: Peter Schuller
>            Priority: Minor
>
> [My aim with this ticket is to explain my current understanding of the FD and it's behavior, express some opinions, and invite others to let me know if I'm misunderstanding something.]
> So I was getting back to CASSANDRA-3569 and I re-read the ticket history, and I want to add a few things that are more about the FD in general, that I've realized since the last batch of discussion.
> Firstly, as a result of investigating gossip more, and of reading CASSANDRA-2597 (Paul Cannon's excellent break-down of what the FD actually does mathematically - thank you Paul!), I now have a much better understanding of the behavior of the failure detector than I did before. Unfortunately for the failure detector. Under the premise that the break-down in CASSANDRA-2597 (and the resulting commit to the behavior of Cassandra) is correct, if we ignore all the guassian/normal distribution stuff (I openly admit I lack the necessary math background for a rigorous analysis), the behavior of the failure detector is the following (not a quote despite use of quote syntax, I'm speaking):
> {quote}
> For a given node, keep track of the last 1000 intervals between heartbeats received via gossip (indirectly or directly from the node, doesn't matter). At any given moment, the phi "score" of the node is the *time since the last heartbeat divided by the average time between heartbeats over the past 1000 intervals* (scaled by some constant factor which is essentially ignoreable, since it is equivalent of the equivalent adjustment to convict threshold). If it goes above a certain value, we consider it down. We check for this on every gossip round (meaning about every second).
> {quote}
> I want to re-state the behavior like this because it makes it trivial to get an intuitive understanding of what it does without dwelwing into the AFD paper.
> In addition, consider that we the failure detector will *immediately* consider a node up when it receives a new heartbeat and if the node is considered down at the time.
> Further, while the accural FD paper talks about giving a non-binary score, and we do this, we don't actually use it except to trigger a binary up/down flap.
> Given the above, and general background, it seems to me that:
> * In terms of detecting that something goes down, the FD is just barely one step above just slapping a fixed timeout on heartbeats; essentially a timeout scaled relative to average historic latency.
> ** But on the other hand, it's also fuzzed (relative to simple tcp timeouts) due to o variation in gossip propagation time which is almost certainly higher than the variation in network latency between two nodes.
> * The gist of the previous two items is that the FD is really not doing anything advanced/magic or otherwise "opaque".
> * In addition, because any heartbeat from a down:ed node implies that it goes up immediately, the failure detector has very little ability to effectively do something "non-trivial" to deal with partial failures, such as demanding that a flapping node show itself healthy for a while before going back up.
> ** Indeed, as far as I can tell, if a node is slowly growing worse in heartbeat it might never get marked as down - if the rate of worsening is slow enough you'll just slowly scale the past latency history and never hit the threshold. (Untested/unconfirmed)
> * The FD is oblivious to input from real traffic (2000 000 messages backed up? doesn't matter, it's "up", even though neighboring nodes have nothing backed up). This is not necessarily wrong in any way, but it needs to be kept in mind when considering what to do *with* the FD.
> Now, CASSANDRA-3910 was recently filed where there is an attempt to use the FD for what I personally think is better dealt with in the request path (see that ticket).
> In seems to me that the FD as it works now is *definitely* not suitable to handle partial failures or smoothly redirecting traffic from anywhere, since we are converting the output of the underlying FD algo to a binary up/down state. Further even if we directly propagated current phi and used it to do relative weighting on sending requests, we still have the instant return to low phi on the very next heartbeat. It is just not suitable, as currently implemented, for anything other than binary up/down flagging as far as I can tell.
> This re-enforces, in my view, my skepticism towards CASSANDRA-2433, and my position in CASSANDRA-3294. We seem to be treating the failure detector as if it's doing something non-trivial, where in reality it isn't (gossip as a communication channel may be non-trivial, but the FD algorithm isn't). It seems to me the main function of the failure detector is that it allows us to scale to very large clusters; we need not to full-mesh ping-pong (whether at app level or at socket level) in order for up/down state to be communicated. This is a useful feature. However, it is actually a result of using gossip as the underlying channel, rather than due to anything specific to the FD algorithm (except insofar as the FD algorithm doesn't need more detailed or specific information than is reasonable to communicate over gossip).
> I believe that the FD *is* good at:
> * Efficiently (in terms of scaling to large clusters) allowing controlled shutdowns and "official" up/down marking.
> ** Useful for e.g. selecting hosts for streaming - but only for it's usefulness as "official flag" marking, not in the sense that it detects failure conditions.
> ** Nodes coming up from scratch can get a feel for the cluster immediately without having to go through an initial burst of failed messages (but currently we don't actually guarantee this anyway because we don't wait for gossip to settle on start-up - but that's a separate issue).
> I believe that the FD is *not* good at:
> * Indicating whether to send a message to a node, or how often, except for the special case of "the node is completely down, don't even bother at all, ever".
> * Optimizing for request latency, at all. It is not an effective tool to mitigate request latencies.
> * Optimizing for avoiding heap growth as requests back up; this is a very separate concern that should take into account things like relative queue sizes, past history of real requests, etc.
> * High latencies, node hiccups, congestion.
> I think that the things it is good at, is a legitimate job for the FD. The things I list under bad, I think is the job of other things like CASSANDRA-2540 and CASSANDRA-3294 to add intelligence to the way we handle requests. Even an infinitely smart FD will never work for these tasks, as long as we retain the binary up/down output (unless you want to start talking flapping up/down at high frequency. I also think that with a sufficiently good request path, we probably don't even need any active ping-pong at all, except maybe very seldom if we don't send any traffic to the node. So while using gossip for this is a cool application of it, I am unconvinced that we need active ping/pong to any great extent if we are sufficiently good at not routing requests to nodes of unknown status, and instantly prefer nodes that are actively responding to those who don't (e.g. least pending request input to routing).
> Finally, there is one additional particularly interesting feature of the failure detector and its integration with gossip the way it works now that I am aware of and feel is worth mentioning: My understanding is that the intention is for the FD, due to coupling with gossip, to guarantee that we never send messages to a node whose impact on the token ring is somehow "incorrect". So for example, on a cluster partition you're supposed to be able to make topological changes during a partition and it will just magically resolve itself when you de-partition. (I would *personally* never ever trust this in production; others may disagree. But theoretically, if it works, it's a cool feature.)
> Thoughts?

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira