[
https://issues.apache.org/jira/browse/ZOOKEEPER-2619?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15599814#comment-15599814
]
Flavio Junqueira commented on ZOOKEEPER-2619:
---------------------------------------------
bq. i think it's a good idea to document this issue in this jira.
Documenting is a good suggestion, but probably better on the official or docs
or at least wiki. I think that's what you're suggesting, yes?
bq. it sounds to me that getConnection() and reenableOps() are basically the
same. right?
I need to step back to respond to this and say how I have interpreted the
original proposal. The description says:
bq. Add a method ZooKeeper.getConnection() returning a ZKConnection object.
ZKConnection would wrap a TCP connection.
My interpretation here is that in {{SendThread}}, instead of starting a
connection automatically once the connection is declared lost, we need to
change the logic to have the application enabling the reconnection. If we let
the application do it and the application takes long, then the session might
end up expiring. I'm not convinced that it is a good idea to let the
application take control over the reconnection.
The different approach with {{reenableOps}} consists of leaving the automatic
reconnection as we have today, but in the {{queuePacket}} method where we have
this:
{code}
if (!state.isAlive() || closing) {
conLossPacket(packet);
} else {
// If the client is asking to close the session then
// mark as closing
if (h.getType() == OpCode.closeSession) {
closing = true;
}
outgoingQueue.add(packet);
}
{code}
we change the predicate of the {{if}} block to evaluate to true also in the
case operations haven't been re-enabled after a connection loss. We might want
to have a new state that represents the client being connected but disabled and
possibly a parameter that disables this behavior to provide what we have
currently for compatibility. I'd say that the default behavior is what we have
today, but I don't feel strongly about it.
bq. one problem i see with reenableOps is that it affects everything using the
zookeeper handle, not just the ops in question.
I'm not sure what you have in mind when you refer to *everything*, [~breed]. If
you're alluding to the fact that it also affects the sync api, then the
configuration switch enables turning it on and off. The discussion below
provides a different option, though.
bq. i kind of like the idea of getting a object that only has async methods
that you can have a strong guarantee of FIFO execution.
I agree, it shouldn't be difficult to separate out so that we have a
synchronous handle and an asynchronous handle. We would need to keep a
{{ZooKeeper}} class that instantiates the two for backward compatibility.
> Client library reconnecting breaks FIFO client order
> ----------------------------------------------------
>
> Key: ZOOKEEPER-2619
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2619
> Project: ZooKeeper
> Issue Type: Bug
> Reporter: Diego Ongaro
>
> According to the USENIX ATC 2010
> [paper|https://www.usenix.org/conference/usenix-atc-10/zookeeper-wait-free-coordination-internet-scale-systems],
> ZooKeeper provides "FIFO client order: all requests from a given client are
> executed in the order that they were sent by the client." I believe
> applications written using the Java client library are unable to rely on this
> guarantee, and any current application that does so is broken. Other client
> libraries are also likely to be affected.
> Consider this application, which is simplified from the algorithm described
> on Page 4 (right column) of the paper:
> {code}
> zk = new ZooKeeper(...)
> zk.createAsync("/data-23857", "...", callback)
> zk.createSync("/pointer", "/data-23857")
> {code}
> Assume an empty ZooKeeper database to begin with and no other writers.
> Applying the above definition, if the ZooKeeper database contains /pointer,
> it must also contain /data-23857.
> Now consider this series of unfortunate events:
> {code}
> zk = new ZooKeeper(...)
> // The library establishes a TCP connection.
> zk.createAsync("/data-23857", "...", callback)
> // The library/kernel closes the TCP connection because it times out, and
> // the create of /data-23857 is doomed to fail with ConnectionLoss. Suppose
> // that it never reaches the server.
> // The library establishes a new TCP connection.
> zk.createSync("/pointer", "/data-23857")
> // The create of /pointer succeeds.
> {code}
> That's the problem: subsequent operations get assigned to the new connection
> and succeed, while earlier operations fail.
> In general, I believe it's impossible to have a system with the following
> three properties:
> # FIFO client order for asynchronous operations,
> # Failing operations when connections are lost, AND
> # Transparently reconnecting when connections are lost.
> To argue this, consider an application that issues a series of pipelined
> operations, then upon noticing a connection loss, issues a series of recovery
> operations, repeating the recovery procedure as necessary. If a pipelined
> operation fails, all subsequent operations in the pipeline must also fail.
> Yet the client must also carry on eventually: the recovery operations cannot
> be trivially failed forever. Unfortunately, the client library does not know
> where the pipelined operations end and the recovery operations begin. At the
> time of a connection loss, subsequent pipelined operations may or may not be
> queued in the library; others might be upcoming in the application thread. If
> the library re-establishes a connection too early, it will send pipelined
> operations out of FIFO client order.
> I considered a possible workaround of having the client diligently check its
> callbacks and watchers for connection loss events, and do its best to stop
> the subsequent pipelined operations at the first sign of a connection loss.
> In addition to being a large burden for the application, this does not solve
> the problem all the time. In particular, if the callback thread is delayed
> significantly (as can happen due to excessive computation or scheduling
> hiccups), the application may not learn about the connection loss event until
> after the connection has been re-established and after dependent pipelined
> operations have already been transmitted over the new connection.
> I suggest the following API changes to fix the problem:
> - Add a method ZooKeeper.getConnection() returning a ZKConnection object.
> ZKConnection would wrap a TCP connection. It would include all synchronous
> and asynchronous operations currently defined on the ZooKeeper class. Upon a
> connection loss on a ZKConnection, all subsequent operations on the same
> ZKConnection would return a Connection Loss error. Upon noticing, the client
> would need to call ZooKeeper.getConnection() again to get a working
> ZKConnection object, and it would execute its recovery procedure on this new
> connection.
> - Deprecate all asynchronous methods on the ZooKeeper object. These are
> unsafe to use if the caller assumes they're getting FIFO client order.
> - No changes to the protocols or servers are required.
> I recognize this could cause a lot of code churn for both ZooKeeper and
> projects that use it. On the other hand, the existing asynchronous calls in
> applications should now be audited anyhow.
> The code affected by this issue may be difficult to contain:
> - It likely affects all ZooKeeper client libraries that provide both
> asynchronous operations and transparent reconnection. That's probably all
> versions of the official Java client library, as well as most other client
> libraries.
> - It affects all applications using those libraries that depend on the FIFO
> client order of asynchronous operations. I don't know how common that is, but
> the paper implies that FIFO client order is important.
> - Fortunately, the issue can only manifest itself when connections are lost
> and transparently reestablished. In practice, it may also require a long
> pipeline or a significant delay in the application thread while the library
> establishes a new connection.
> - In case you're wondering, this issue occurred to me while working on a new
> client library for Go. I haven't seen this issue in the wild, but I was able
> to produce it locally by placing sleep statements in a Java program and
> closing its TCP connections.
> I'm new to this community, so I'm looking forward to the discussion. Let me
> know if I can clarify any of the above.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
