Dear Ariel,

1) Profiling

WP-MIRROR 0.6 saw the introduction of a `--profile' command-line option. Which provides a detailed breakdown of where time is spent during a mirror build. Unfortunately, WP-MIRROR 0.5 did not have this feature, so only aggregate comparisons are possible.

2) Performance studies

Much of the Winter and Spring was devoted to time trials. These are all documented in WP-MIRROR 0.6 Reference Manual, Appendix G. The following may be very interesting for you

2.1) G.8 Experiments with InnoDB data compression

Currently, InnoDB offers two on-dist storage formats, Antelope and Barracuda. The later offers data compress. I performed experiments on many of the largest wiki tables (e.g. categorylinks, image, langlinks, pagelinks, templatelinks, text) to determine the space savings and time penalty. For a summary of results, please take a look at Table G.1 Database table size v. ROW_FORMAT and KEY_BLOCK_SIZE.

2.2) G.9 Experiments with commonswiki.image

For several reasons, I wanted to import commonswiki.image as part of the mirror build process. That table is large and takes time to import. However, once imported WP-MIRROR 0.6 saves a greater amount of time elsewhere (e.g. I do not have to scrape the XML dumps for image file names).

Because of the size of commonswiki.image, I performed a lengthy series of experiments to determine the fastest way of importing it. Many of the best methods make use of features first offered in MySQL 5.5 and InnoDB 1.1 (e.g fast index creation). For a summary of results, please take a look at Figure G.1 InnoDB Experiments Importing commonswiki.image Table.

3) Documentation

The above mentioned WP-MIRROR 0.6 Reference Manual may be found at http://www.nongnu.org/wp-mirror/manual/. It is also included in with the DEB package.

4) Questions

4.1) mwxml2sql

While reading the code for `mwxml2sql.c', I noticed `-n, --nodrop' option which sets KEY_BLOCK_SIZE=16 for the `text' table. From my experiments (see Appendix G.8) I do not think that any compression takes place. KEY_BLOCK_SIZE=4 might be a better choice.

4.2) INSERT IGNORE vs. REPLACE INTO

For the initial mirror build, INSERT INTO commands get the job done. However, for updates to an existing mirror, I would prefer not to DROP TABLE every time. For this reason, I rewrite the INSERT INTO commands into REPLACE INTO commands. This rewrite works fine.

What I would like to know from you is this: for which table is INSERT IGNORE a better choice than REPLACE INTO. I can see that the `revision' and `text' table are candidates because they are never updated, only added to. But what about the other tables? Any advise would be appreciated.

Sincerely Yours, Kent

On 6/3/13, Ariel T. Glenn ariel@wikimedia.org wrote:

Στις 03-06-2013, ημέρα Δευ, και ώρα 10:22 -0400, ο/η wp mirror έγραψε:

...

Dear list members,

I am pleased to announce the release of WP-MIRROR 0.6.

The main design objective was this: PERFORMANCE. WP-MIRROR 0.6 now builds the `enwiki' (which is the most demanding case) with 80% less time and 75% less memory than v0.5.

Sounds great! Can you give us some benchmarks? I'm particularly interested in length of time with the old and nw versions of your package for the various stages of setting up a dump of current pages for the English language Wikipedia, on whatever hardware you are using for testing.

Ariel