[MarkLogic Dev General] "Joins" in search:search or cts:search
Lee, David
dlee at epocrates.com
Fri Nov 18 10:40:39 PST 2011
Thanks.
On my system (a "large" EC2 instance on EBS) I'm seeing about 5,000 document deletes/sec
When i'm trying to delete 5 mil documents that still goes over timeout periods so I have to do fancy dancing to get it to works.
----------------------------------------
David A. Lee
Senior Principal Software Engineer
Epocrates, Inc.
dlee at epocrates.com<mailto:dlee at epocrates.com>
812-482-5224
From: general-bounces at developer.marklogic.com [mailto:general-bounces at developer.marklogic.com] On Behalf Of Aaron Redalen
Sent: Friday, November 18, 2011 12:53 PM
To: General MarkLogic Developer Discussion
Subject: Re: [MarkLogic Dev General] "Joins" in search:search or cts:search
Just as a side note, collection-delete and directory-delete are able to run in "fast" mode if the following conditions are met:
* The database config parameter for "directory-creation" must be set to manual
* No triggers
* No auditing of updates
* There should be no lock fragments at the time of the call to collection-delete (those established via xdmp:lock-acquire)
If any of these conditions are not met, collection-delete must retrieve and delete each document, which may take awhile for large collections.
However, if these conditions are met, collection-delete doesn't need to retrieve any fragments. Instead, it simply sets the deleted timestamp on the documents matching the collection. In this mode, we can delete tens of thousands of documents per second.
Aaron Redalen
Director, Professional Services
MarkLogic Corporation
aaron.redalen at marklogic.com<mailto:aaron.redalen at marklogic.com>
Phone: +1 650 655-2349
Cell: +1 240 688-7433
www.marklogic.com<http://www.marklogic.com>
From: "Lee, David" <dlee at epocrates.com<mailto:dlee at epocrates.com>>
Reply-To: General MarkLogic Developer Discussion <general at developer.marklogic.com<mailto:general at developer.marklogic.com>>
Date: Thu, 17 Nov 2011 13:57:28 -0800
To: General MarkLogic Developer Discussion <general at developer.marklogic.com<mailto:general at developer.marklogic.com>>
Subject: Re: [MarkLogic Dev General] "Joins" in search:search or cts:search
Thanks, no I have not measured the speed of collection-delete() of a non-existant collection.
But I *have* timed collection-delete() which contains millions of documents and its exceeds request times ... so I have to split it up.I wanted to avoid starting 100,000 spawned threads of which 5 actually did anything.
----------------------------------------
David A. Lee
Senior Principal Software Engineer
Epocrates, Inc.
dlee at epocrates.com<mailto:dlee at epocrates.com>
812-482-5224
From: general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com> [mailto:general-bounces at developer.marklogic.com] On Behalf Of Geert Josten
Sent: Thursday, November 17, 2011 3:24 PM
To: General MarkLogic Developer Discussion
Subject: Re: [MarkLogic Dev General] "Joins" in search:search or cts:search
Not index-fast no, true, but you're not retrieving data from the database, so saves a lot of fuss..
Have you tried measuring the speed of a collection-delete of an empty collection? Interesting case.. :)
Kind regards,
Geert
Van:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com> [mailto:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com>] Namens Lee, David
Verzonden: donderdag 17 november 2011 21:17
Aan: General MarkLogic Developer Discussion
Onderwerp: Re: [MarkLogic Dev General] "Joins" in search:search or cts:search
Deletes in ML have not been "fast" by any metric in my experience.
I'll try just using the estimate ... maybe thats faster.
----------------------------------------
David A. Lee
Senior Principal Software Engineer
Epocrates, Inc.
dlee at epocrates.com<mailto:dlee at epocrates.com>
812-482-5224
From:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com> [mailto:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com>] On Behalf Of Geert Josten
Sent: Thursday, November 17, 2011 2:54 PM
To: General MarkLogic Developer Discussion
Subject: Re: [MarkLogic Dev General] "Joins" in search:search or cts:search
Hi Lee,
Actually, the exists() in your code might be the slowest part of your code. The collection call is likely backed by an index, so quick. The exists works on a sequence however. It could be that it is optimized under the hood to use xdmp:estimate in this case, but not sure. Could try to rewrite that. But actually, I would test at all.
A collection-delete of an empty collection won't take time I'd say. So wouldn't worry about that too much.
What remains is the initial collection, which returns a sequence. If you are not collecting the results, MarkLogic doesn't need to keep it in memory. Could very well be that it is streamed in the outer for loop. Otherwise try chunking it in batches of 10k. Remember that deletes in ML are fast! It's just a flag on each fragment..
Kind regards,
Geert
Van:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com> [mailto:general-bounces at developer.marklogic.com<mailto:general-bounces at developer.marklogic.com>] Namens Lee, David
Verzonden: donderdag 17 november 2011 20:41
Aan: General Mark Logic Developer Discussion (general at developer.marklogic.com<mailto:general at developer.marklogic.com>)
Onderwerp: [MarkLogic Dev General] "Joins" in search:search or cts:search
I suspect the answer is "no" ... but just plugging the brains out there ..
For good or bad I use this architype.
I have many "summary" documents say "/logs/1.xml" , "/logs/2.xml" which belongs to the collection "/summaries"
There can be many (100k+)
Each summary document lists a refernce to external URL's (in this case Amazon S3) from which data could be loaded.
If I load the data I put each group into a collection named by the URL of the summary.
So say I have 10,000 XML documents referenced by doc("/logs/1.xml") If I choose to load them, they will end up in collection
"/logs/1.xml". These summaries are in the collection say "/summaries"
The reason for this is for the ability to easily bulk delete blocks of documents based on their summaries.
I can list the summaries and by a simple
exists( collection( $url) )
cant tell if any actual log documents have been loaded.
NOW: I want to be able to delete all records by summary but only if the documents have been loaded.
Suppose I had 100k summary URL's I could do
for $url in collection("/summaries")
if( exists( collection( $url) ) then
xdmp:collection-delete($url)
else ()
This works and all ... but suppose I want something more efficiient.
Overall there may be only say 1% of the summary documents actually loaded. Furthermore if there were LOTS of ones loaded the above would timeout.
So I spawn a thread to delete say [1 to 10] of every summary collection ...
but say I have 100k collections most of the threads do nothing.
So I have to revert to the above to first check if the collection has anything before spawning a thread.
Quesiton: Is there a cts:search option which can do a collection query based on the results of the search itself ?
that is (pseudo code)
in one cts:search
for $c in collection("x")/document-uri(.)
if( exists( collection( $c) )
return $c
doing this in FLOWR is very slow ...
but its what I'm resorting to ....
----------------------------------------
David A. Lee
Senior Principal Software Engineer
Epocrates, Inc.
dlee at epocrates.com<mailto:dlee at epocrates.com>
812-482-5224
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://developer.marklogic.com/pipermail/general/attachments/20111118/29ae60ef/attachment-0001.html
More information about the General
mailing list