atom feed2 messages in com.marklogic.developer.generalRe: [MarkLogic Dev General] [RESOLVE]...
FromSent OnAttachments
Eliot KimberNov 10, 2017 1:44 pm 
Ron HitchensNov 10, 2017 5:44 pm 
Subject:Re: [MarkLogic Dev General] [RESOLVE] Spawned Task Appears to Block Other Threads
From:Ron Hitchens (ro@ronsoft.com)
Date:Nov 10, 2017 5:44:53 pm
List:com.marklogic.developer.general

I posted a few messages about this a couple of months ago with reasonably in-depth discussion of locks, deadlocks and concurrency.

To really gain a solid understanding of how this stuff works, get to know and love the most excellent paper *Inside MarkLogic Server*, available here: https://developer.marklogic.com/inside-marklogic

---- Ron Hitchens ro@overstory.co.uk, +44 7879 358212

On November 10, 2017 at 9:44:41 PM, Eliot Kimber (ekim@contrext.com) wrote:

I moved my response handler to a different web app than the web app that starts the job submission task, but I was still having the issue where the response handler didn’t handle responses (or didn’t get responses, hard for me to tell which it is) until the job submission task is canceled.

I realized that my problem was the initial job submission was updating the job record for each job, but I was doing the update as part of the main processing, rather than using eval(), so the commit wasn’t done until the task ended.

The response handler also wants to update the job record, but because there is a pending commit, it is blocked.

By having the job submission do the update via eval() then the commit is done immediately and everything works.

So the lesson is: you have to really understand the implications of updates and commits or you will go astray. Or maybe “concurrency is always hard”.

Thanks,

Eliot

-- Eliot Kimber http://contrext.com

On 11/10/17, 9:39 AM, "gene@developer.marklogic.com on behalf of Eliot Kimber" <gene@developer.marklogic.com on behalf of ekim@contrext.com> wrote:

Yes, the client process is started from a Web app, so I think your analysis is correct.

I will move the response handling to a separate Web app—probably should have done that from the start.

Thanks,

Eliot

-- Eliot Kimber http://contrext.com

On 11/9/17, 11:46 PM, "gene@developer.marklogic.com on behalf of Geert Josten" <gene@developer.marklogic.com on behalf of Geer@marklogic.com> wrote:

Hi Eliot,

I think you kicked off your watcher job with an HTTP request, and it keeps the port open until it finishes. Only one thread can use the port at the same time. Use a different port for task response traffic, or consider running your watcher as a scheduled task.

Not super robust, and probably not used in production, but i did write an alternative queque for MarkLogic. It might give you some ideas..

https://github.com/grtjn/ml-queue

Cheers, Geert

On 11/10/17, 1:06 AM, "gene@developer.marklogic.com on behalf of Eliot Kimber" <gene@developer.marklogic.com on behalf of ekim@contrext.com> wrote:

I have a system where I have a ³client² ML server that submits jobs to a set of remote ML servers, checking their task queues and keeping each server¹s queue at a max of 100 queued items (the remote servers could go away without notice so the client needs to be able to restart tasks and not have too many things queued up that would just have to resubmitted).

The remote tasks then talk back to the client to report status and return their final results.

My job submission code use recursive functions to iterate over the set of tasks to be submitted, checking for free remote queue slots via the ML REST API and submitting jobs as the queues empty. This code is spawned into a separate task in the task server. It uses xdmp:sleep(1000) to pause between checking the job queues.

This all works fine, in that my jobs are submitted correctly and the remote queues fill up.

However, as long as the job-submission task in the task server is running, the HTTP app that handles the REST calls from the remote servers is blocked (which blocks the remote jobs, which are of course waiting for responses from the client).

If I kill the task server task, then the remote responses are handled as I would expect.

My question: Why would the task server task block the other app? There must be something I¹m doing or not doing but I have no idea what it might be.