lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Robert Muir (JIRA)" <j...@apache.org>
Subject [jira] Commented: (LUCENE-2905) Sep codec writes insane amounts of skip data
Date Fri, 04 Feb 2011 18:18:30 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-2905?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12990662#comment-12990662
] 

Robert Muir commented on LUCENE-2905:
-------------------------------------

Renaud thanks for the paper... I will spend some time trying to digest it!
But I think its always an option to try to reduce the number of files, too.
This is important also for # of open files and other practical reasons.

Mike a few questions:
bq. If we are REALLY sure keeping int alignment in these intblock
encoded files is not important (ie, we really do get best perf by
slurping in byte[] and then decoding from there), then we should also
store eg skip data into the frq/doc file (this is what Standard
does).

Well, I measured this a lot, but why box ourselves out? As a first step 
we can still keep the .skp file as-is, but it only needs point to 
the start of the doc block in the frq/doc file.

bq. Maybe similarly interleave payload/positions packets?

I think we should do something here. But when i started trying to draw
this up, I came to the conclusion that payload byte[]s should themselves
be actual terms (e.g. deduplicated), and we store some sort of ord to get
to them instead of bytes and length, etc. if they were themselves terms,
then they could also store forward postings back to their docs, and you 
could query on payloads (attributes) efficiently too... but I know this
would be a fairly large change.

bq. Separately, I think we should break out "when skip is even
stored" vs "how frequently we index skip data".

I agree, another reason to pull skipInterval completely codec-private.
Then a codec could itself have a separate "skipMinimum" too.

bq. For low DF terms w/in a block I think we shouldn't store their
pointers into the posting; instead, you should load an earlier term's
postings and scan over its postings. This should save tons of space
in the tib file.

How would this work? Isnt everything right now in the .tib delta-encoded
against the index term? What if there are 'large' terms in between?
And for some queries like rangequery, wouldnt this create a little O(n^2)
of sorts? I don't think this is a big deal, most people should be using
e.g. NumericRangeQuery, and maybe we could still prevent it...?


> Sep codec writes insane amounts of skip data
> --------------------------------------------
>
>                 Key: LUCENE-2905
>                 URL: https://issues.apache.org/jira/browse/LUCENE-2905
>             Project: Lucene - Java
>          Issue Type: Bug
>            Reporter: Robert Muir
>             Fix For: Bulk Postings branch
>
>
> Currently, even if we use better compression algorithms via Fixed or Variable Intblock
> encodings, we have problems with both performance and index size versus StandardCodec.
> Consider the following numbers:
> {noformat}
> standard:
> frq: 1,862,174,204 bytes
> prx: 1,146,898,936 bytes
> tib: 541,128,354 bytes
> complete index: 4,321,032,720 bytes
> bulkvint:
> doc: 1,297,215,588 bytes
> frq: 725,060,776 bytes
> pos: 1,163,335,609 bytes
> tib: 729,019,637 bytes
> complete index: 5,180,088,695 bytes
> simple64:
> doc: 1,260,869,240 bytes
> frq: 234,491,576 bytes
> pos: 1,055,024,224 bytes
> skp: 473,293,042 bytes
> tib: 725,928,817 bytes
> complete index: 4,520,488,986 bytes
> {noformat}
> I think there are several reasons for this:
> * Splitting into separate files (e.g. postings into .doc + .freq). 
> * Having to store both a relative delta to the block start, and an offset into the block.
> * In a lot of cases various numbers involved are larger than they should be: e.g. they
are file pointer deltas, but blocksize is fixed...
> Here are some ideas (some are probably stupid) of things we could do to try to fix this:
> Is Sep really necessary? Instead should we make an alternative to Sep, Interleaved? that
interleaves doc and freq blocks (doc,freq,doc,freq) into one file? the concrete impl could
implement skipBlock() for when they only want docdeltas: e.g. for Simple64 blocks on disk
are fixed size so it could just skip N bytes. Fixed Int Block codecs like PFOR and BulkVint
just read their single numBytes header they already have today, and skip numBytes.
> Isn't our skipInterval too low? Most of our codecs are using block sizes such as 64 or
128, so a skipInterval of 16 seems a little overkill.
> Shouldn't skipInterval not even be a final constant in SegmentWriteState, but instead
completely private to the codec?
> For block codecs, doesn't it make sense for them to only support skipping to the start
of a block? Then, their skip pointers dont need to be a combination of delta + upto, because
upto is always zero. What would we have to modify in the bulkpostings api for jump() to work
with this?
> For block codecs, shouldn't skipInterval then be some sort of divisor, based on block
size (maybe by default its 1, meaning we can skip to the start of a every block)
> For codecs like Simple64 that encode fixed length frames, shouldnt we use 'blockid' instead
of file pointer so that we get smaller numbers? e.g. simple64 can do blockid * 8 to get to
the file pointer.
> Going along with the blockid concept, couldnt pointers in the terms dict be blockid deltas
from the index term, instead of fp deltas? This would be smaller numbers and we could compress
this metadata better.

-- 
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message