[
https://issues.apache.org/jira/browse/ZOOKEEPER-2471?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16136834#comment-16136834
]
Dan Benediktson commented on ZOOKEEPER-2471:
--------------------------------------------
Thanks for the reviews, [~nickt] and [~hanm]. Sorry, I should have been more
clear about the scenario: although I didn't realize it at the time I initially
filed this bug (I wasn't aware of the extent of the divergence of my own fork
from mainline), I don't think you can reasonably run into an infinite connect
loop on *current* Apache Zookeeper. On the current code base, hitting an
infinite connect loop basically requires falling victim to the random number
generator repeatedly; not likely at all.
However, if you pick up exponential backup from ZOOKEEPER-2869, it becomes a
certainty if you hit the right conditions: basically, you just need to get
enough initial failures to rack up a large exponential sleep backoff, which in
practice happened to us during leader election on a large ensemble, and then
your sleep time can potentially outgrow your connect timeout. That's somewhat
dependent on the backoff algorithm used, but the one we have is pretty
reasonable in how it caps the backoff, and it was certainly still possible
there: we had applications which went down completely since they were unable to
connect to ZK until we recycled the application every time we leader elected
the ensemble they talked to, until we fixed this.
So, that's why I suggested in the pull request that maybe a compromise would be
to submit tests along with ZOOKEEPER-2869: a lot of the machinery that's needed
to make this testable is actually stuff that can be introduced as part of the
product feature in ZOOKEEPER-2869.
> Java Zookeeper Client incorrectly considers time spent sleeping as time spent
> connecting, potentially resulting in infinite reconnect loop
> ------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: ZOOKEEPER-2471
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2471
> Project: ZooKeeper
> Issue Type: Bug
> Components: java client
> Affects Versions: 3.5.3
> Environment: all
> Reporter: Dan Benediktson
> Assignee: Dan Benediktson
> Attachments: ZOOKEEPER-2471.patch
>
>
> ClientCnxnSocket uses a member variable "now" to track the current time, and
> lastSend / lastHeard variables to track socket liveness. Implementations, and
> even ClientCnxn itself, are expected to call both updateNow() to reset "now"
> to System.currentTimeMillis, and then call updateLastSend()/updateLastHeard()
> on IO completions.
> This is a fragile contract, so it's not surprising that there's a bug
> resulting from it: ClientCnxn.SendThread.run() calls updateLastSendAndHeard()
> as soon as startConnect() returns, but it does not call updateNow() first. I
> expect when this was written, either the expectation was that startConnect()
> was an asynchronous operation and that updateNow() would have been called
> very recently, or simply the requirement to call updateNow() was forgotten at
> this point. As far as I can see, this bug has been present since the
> "updateNow" method was first introduced in the distant past. As it turns out,
> since startConnect() calls HostProvider.next(), which can sleep, quite a lot
> of time can pass, leaving a big gap between "now" and now.
> If you are using very short session timeouts (one of our ZK ensembles has
> many clients using a 1-second timeout), this is potentially disastrous,
> because the sleep time may exceed the connection timeout itself, which can
> potentially result in the Java client being stuck in a perpetual reconnect
> loop. The exact code path it goes through in this case is complicated,
> because there has to be a previously-closed socket still waiting in the
> selector (otherwise, the first timeout evaluation will not fail because "now"
> still hasn't been updated, and then the actual connect timeout will be
> applied in ClientCnxnSocket.doTransport()) so that select() will harvest the
> IO from the previous socket and updateNow(), resulting in the next loop
> through ClientCnxnSocket.SendThread.run() observing the spurious timeout and
> failing. In practice it does happen to us fairly frequently; we only got to
> the bottom of the bug yesterday. Worse, when it does happen, the Zookeeper
> client object is rendered unusable: it's stuck in a perpetual reconnect loop
> where it keeps sleeping, opening a socket, and immediately closing it.
> I have a patch. Rather than calling updateNow() right after startConnect(),
> my fix is to remove the "now" member variable and the updateNow() method
> entirely, and to instead just call System.currentTimeMillis() whenever time
> needs to be evaluated. I realize there is a benefit (aside from a trivial
> micro-optimization not worth worrying about) to having the time be "fixed",
> particularly for truth in the logging: if time is fixed by an updateNow()
> call, then the log for a timeout will still show exactly the same value the
> code reasoned about. However, this benefit is in my opinion not enough to
> merit the fragility of the contract which led to this (for us) highly
> impactful and difficult-to-find bug in the first place.
> I'm currently running ant tests locally against my patch on trunk, and then
> I'll upload it here.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
