Matomo Issue #2243 - Bulk insert reports when archiving

@mattab opened this Issue on March 30th 2011 Member

Extract from Custom date range processing with sql profiler enabled (see global.ini.php)


Executed 5314 times in 5313.7ms (average = 1ms)

    INSERT IGNORE INTO piwik_archive_blob_2010_01 (idarchive, idsite, date1, date2, period, ts_archived, name, value) VALUES (?,?,?,?,?,?,?,?)

What happens is that the Page URL / Page Titles have many children, ie. thousands, and Piwik will INSERT one at a time, in this case 5313 times. Instead, Piwik should log all these into a file and use LOAD DATA INFILE

I'm pretty sure we'll hit some bugs with some user configurations, so plan for a setting in config file to disable the bulk insert.

@mattab commented on March 31st 2011 Member

Attachment:
unhex test.txt

@robocoder commented on April 5th 2011 Contributor

Attachment: Hack for charset mismatch
2243.patch

@mattab commented on March 30th 2011 Member

(In [4239]) Refs #2243

better logging
better performance when archiving Custom Date Range by changing cache key
remove unnecessary class hierarchy - ArchiveProcessing/Record*
Adding helper function to load data bulk in mysql + unit tests - tested with Mysql PDO but not mysqli (are webtests still running with mysqli ?)
If the bulk import fails for any reason, we fallback to loading results in a loop and doing plain INSERTs (rather than config file to disable)

@robocoder commented on March 30th 2011 Contributor

are webtests still running with mysqli

Only unit tests. I'll address this when we have parameterized builds running on Jenkins.

@mattab commented on March 30th 2011 Member

OK cool I actually meant unit tests, that's perfect

@mattab commented on March 31st 2011 Member

(In [4253]) Fixes #2243

Enabling the bulk loading on BLOB records.
Had to first run the blobs through bin2hex... indeed even though it looked like it was working at first, the tests once again saved us from a VERY nasty bug which would have shown up only later and be extremely difficult to find. LOAD DATA INFILE was working fine most of the time, but a few tests were failing! Then I discovered that the file containing the blob wasn't loaded properly in all cases, unless the binary blob was hex'ed.
I ran quick performance test and it seems approximately the same performance, and uses the same disk space as well.
However in some cases, it will definitely improve performance.
memory usage somehow is much less in my test! so definitely a great point.
it would have been nice to have versionning in the archive_blob table, since now we have 2 different data types (old blob are pure binary, new blobs are hex'ed) but it is working nonetheless (Old blobs will fail the first gzuncompress(hex2bin()) then fallback to gzuncompress() only

@robocoder commented on March 31st 2011 Contributor

Can you use MySQL's native HEX() and UNHEX() functions? I would think this would give us backward compatibility on blobs while also reducing php's runtime memory usage (because the string conversion happens in the db server).

@mattab commented on March 31st 2011 Member

To use HEX we would have to use LOAD DATA INFILE ....... SET value = HEX(value) but the SET feature is only available as of 5.0.3

For the select part, I gave it a quick try, but it doesn't seem to work. Attached patch for reference.
in any case if this code is working perfectly as I hope it does, then it's fine ;)

@robocoder commented on March 31st 2011 Contributor

That wasn't what I meant. Let me test something.

@robocoder commented on March 31st 2011 Contributor

I'm still investigating an alternative to hex2bin/bin2hex because the current method doubles the blob size, but I want to log some observations first:

re: r4259. Not sure why it passed (before commenting it out) on your box. The CI server runs 5.3.6, so it's falling back to databaseInsertIterate. I suspect the test fails here because it inserts binary data (e.g., containing NULs) into a varchar column.

MySQL documentation says LOAD DATA INFILE uses the database's character set but we haven't explicitly set this in the CREATE DATABASE. Also, the "CHARACTER SET" clause wasn't added to the LOAD DATA INFILE syntax until MySQL 5.0.38. (At minimum, we should set the database charset for consistency.)

One thought is conditionals to use the newer features if available, e.g., Zend_Db_Adapters have a getServerVersion() method.

@robocoder commented on March 31st 2011 Contributor

(In [4263]) refs #2243 - set database's default charset explicitly

@robocoder commented on March 31st 2011 Contributor

(In [4264]) refs #2243 - use Piwik_Exec() for MYSQLI compatibility, i.e., can't prepare LOAD DATA INFILE when using Piwik_Query(). Tested with 5.1.3, 5.1.6, 5.2.0, and 5.2.17.

@robocoder commented on March 31st 2011 Contributor

I'll add a test case that uses a table with a blob column It appears the bin2hex/hex2bin is unnecessary, so long as the column's datatype is blob.

@mattab commented on March 31st 2011 Member

You says it doubles the blob size, but my tests shows that in fact it doesn't.
I run the archiving on 2 full years with the batch insert, and wihout, and the DB size was the same in both cases.

OK for charset, good thinking.

I'm just slightly concerned about the added escaping function, but tests are passing which should be fine.

When you test, it is good to also not only rely on unit tests: the blob sizes & scope they generate is quite limited. Try to generate loads of data using Visito Generator, then archive the data with the old & new algorithm to see if both work as excpected. You can also easily compare piwikarchive* db size.
To switch from one to the other, simply do:

### Eclipse Workspace Patch 1.0
#P trunk
Index: core/ArchiveProcessing.php
===================================================================
--- core/ArchiveProcessing.php  (revision 4260)
+++ core/ArchiveProcessing.php  (working copy)
@@ -696,7 +696,7 @@
    protected function insertBulkRecords($records)
    {
        // Using standard plain INSERT if there is only one record to insert
-       if($DEBUG_DO_NOT_USE_BULK_INSERT = false
+       if($DEBUG_DO_NOT_USE_BULK_INSERT = !false
            || count($records) == 1)
        {
            foreach($records as $record)

@robocoder commented on April 1st 2011 Contributor

But that doesn't disable the bin2hex/hex2bin. When I say "double the size", I'm comparing the old blobs (octets) vs the new blobs (nybbles as hexits). So, DEBUG_DO_NOT_USE_BULK_INSERT doesn't affect the size as it still uses hexits.

With the escaping function (used only when writing to the infile), blobs go back to being binary data.

@robocoder commented on April 1st 2011 Contributor

(In [4272]) refs #2243 - looks good so far

@mattab commented on April 1st 2011 Member

But that doesn't disable the bin2hex/hex2bin.

oh oh I messed up! and I thought that the results were strange, well I should have thought a bit more indeed

The tests are failing for missing table: http://qa.piwik.org:8080/webtest/003_UnitTests/001_response_invoke.html

when you are confident all is working and tested properly, please ping me and I will test on the demo :)

@robocoder commented on April 1st 2011 Contributor

Yeah, I forgot to drop the database before running my unit test changes. I'll fix this a.m.

@robocoder commented on April 1st 2011 Contributor

Fixing the unit tests is taking longer than I thought ... encountering more test runner side-effects (similar to #896).

@robocoder commented on April 1st 2011 Contributor

(In [4285]) refs #2243 - move test code around