accumulo-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Christopher Tubbs (JIRA)" <>
Subject [jira] [Commented] (ACCUMULO-775) Optimize iterator seek() method when seeking forward
Date Sun, 30 Sep 2012 01:52:08 GMT


Christopher Tubbs commented on ACCUMULO-775:

David, I think we have to first understand why it is faster sometimes to scan instead of seek
(even the fast seek that Adam refers to). I'm not actually sure why it should happen, except
perhaps it's just more expensive to make the decision than it is for users to simply know
their data well enough to guess at a reasonable number of entries over which to scan before

If somebody can simulate this, or demonstrate where the slow-down is, I'd love to know, but
considering that we already have the optimization Adam mentioned, I'm thinking this is probably
a low priority thing to investigate.
> Optimize iterator seek() method when seeking forward
> ----------------------------------------------------
>                 Key: ACCUMULO-775
>                 URL:
>             Project: Accumulo
>          Issue Type: Improvement
>          Components: tserver
>            Reporter: Christopher Tubbs
>            Assignee: Keith Turner
>            Priority: Minor
>              Labels: iterator, scan, seek
>             Fix For: 1.5.0
> At present, seeking is a very expensive operation. Yet, it is a very common case, especially
when writing filtering/consuming/skipping iterators to seek to the next possible match (perhaps
in the next row, when matching a column family with a regular expression), rather than continuing
to iterate. A common solution to the problem of whether to scan or seek is to continue to
scan for some threshold (~10-20 entries), hoping to just "run into" the next possible match,
rather than waste resources seeking directly to it.
> This pattern can be rolled in to the lower level iterator, so that iterators on top don't
have to do this. They can seek, and the underlying source iterator can simply consume the
next X entries when it makes sense, rather than waste resources seeking.
> I could be wrong (please comment and correct me below if I am), but I imagine that the
places where this would make the most sense is if the data currently being sought (seek'd)
is in the current compressed block from the underlying file, especially if it is forward,
relative to the current pointer. A better seek method should be able to tell where one currently
is, and whether the requested data is within reach without doing all the expensive operations
to re-seek to the same compressed block that is already loaded, reload it, decompress it,
and scan to the requested starting point.
> Having such an optimization would eliminate the need for users to try to calibrate their
own such scan vs. seek optimization based on guessing whether their data is in the current
block or another one, while still getting that same performance benefit.

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see:

View raw message