[
https://issues.apache.org/jira/browse/LUCENE-9191?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael McCandless updated LUCENE-9191:
---------------------------------------
Attachment: LUCENE-9191.patch
Status: Open (was: Open)
First cut at this, I think it's close!
I created a simple Python tool, {{dev-tools/scripts/create_line_file_docs.py}},
that downloads Europarl v7 corpus, un-tgzs it into a temp dir, extracts all
body text + titles + dates, breaks documents into "approximately" 1 K
characters (trying to split at next word boundary) accord to clipped normal
distribution, shuffles the resulting large full line docs file, and then spits
out ~20, ~200, ~2000 MB sized files (before compression).
Then, it compresses these files in chunks, saving the valid seek points. It
turns out you can concatenate multiple gzip'd chunks into a single file and the
resulting file is also a valid {{.gz}} file.
Then I fixed {{LineFileDocs.java}} to also load the valid seek points (I stored
in a separate {{.seek}} file side-by-side with the gzip'd line docs file), and
finally, randomly seek to one of those points on init/reopen.
I confirmed Lucene core tests pass ({{./gradlew lucene:core:test}}) on each of
three line doc + seek files. I plan to commit the smallest one, and make the
medium and large ones available at {{home.apache.org}} after pushing.
Note that this is a hard break – after this change, {{LineFileDocs.java}} is no
longer able to skip (scan) slowly through a line docs file, as it does today. I
thought this was safest so people don't accidentally fall into an otherwise
silent performance trap. I think it's intensely unlikely that anyone is using
a different line docs file than what's committed :) I'm probably the only
person on the planet that ever does this (for better test coverage)!
> Fix linefiledocs compression or replace in tests
> ------------------------------------------------
>
> Key: LUCENE-9191
> URL: https://issues.apache.org/jira/browse/LUCENE-9191
> Project: Lucene - Core
> Issue Type: Task
> Reporter: Robert Muir
> Assignee: Michael McCandless
> Priority: Major
> Attachments: LUCENE-9191.patch
>
>
> LineFileDocs(random) is very slow, even to open. It does a very slow "random
> skip" through a gzip compressed file.
> For the analyzers tests, in LUCENE-9186 I simply removed its usage, since
> TestUtil.randomAnalysisString is superior, and fast. But we should address
> other tests using it, since LineFileDocs(random) is slow!
> I think it is also the case that every lucene test has probably tested every
> LineFileDocs line many times now, whereas randomAnalysisString will invent
> new ones.
> Alternatively, we could "fix" LineFileDocs(random), e.g. special compression
> options (in blocks)... deflate supports such stuff. But it would make it even
> hairier than it is now.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
