-Re: a more fault-tolerant collector
Ariel Rabkin 2010-10-12, 17:40
There is support for assuming writes are asynchronous. I don't know if
it's ever been tried in production though. It also doesn't quite
solve the problem.
Right now, collectors only return on success; on a write failure, they exit.
We would need to figure out if a write failure means a 200 OK with a
special body of the response, or a different HTTP response code.
Likewise, agents need to be modified to handle that return code.
And we have a bit of a design question. If a collector is up, but
can't write, should the agent fail over to a new one, or just retry
On Tue, Oct 12, 2010 at 10:25 AM, Bill Graham <[EMAIL PROTECTED]> wrote:
> We had problems with a single DN the other day and all collectors
> ultimately died after trying N failed attempts. I believe at least one
> of the failures was was during a commit.
> I think backpressure sounds like the right approach, bit it seems like
> there would be some practical challenges, particularly around async
> writes or commits to HDFS. Ari, how does this behavior work currently,
> and would this be difficult to handle?
> In oahc.datacolleciton.writer.SeqFileWriter there's an exception block
> with this in it:
> // We don't want to loose anything
> log.fatal("IOException when trying to write a chunk, Collector is
> going to exit!", e);
> isRunning = false;
> On Tue, Oct 12, 2010 at 9:57 AM, Eric Yang <[EMAIL PROTECTED]> wrote:
>> I thought that is what it is currently doing with one twist, the commit and response is async. Collector exits if the file system is unavailable for extensive period of time. If it is not doing what's described above, then we definitely should fix it.
>> On 10/11/10 10:49 PM, "Ariel Rabkin" <[EMAIL PROTECTED]> wrote:
>> This is an answer to a question Bill asked me recently: can we
>> redesign the Collector process to behave better if the filesystem is
>> I think we can do this by backpressure. If the write fails, the
>> collector should return an error to the agent. And the agent should
>> treat it like a post failure, and retry. Thoughts?
>> Ari Rabkin [EMAIL PROTECTED]
>> UC Berkeley Computer Science Department
Ari Rabkin [EMAIL PROTECTED]
UC Berkeley Computer Science Department