Here we take a closer look at the ProbeSetData table and indices. At the end we have reduced the size by 25% and decided to keep the myisam format for this single table. In the future it may be replaced by lmdb or similar.
This is by far the largest table (~200Gb). I need to add disk space to be able to host it on the NVME and move stuff around. Final move is GN2 code and we have over 400Gb free.
This time I failed porting to InnoDB (see Migration below):
I disabled these and they need to be restored:
The following fetches the data to be used for mapping:
SELECT
Strain.Name, ProbeSetData.value, ProbeSetSE.error, NStrain.count, Strain.Name2
FROM
(ProbeSetData, ProbeSetFreeze,
Strain, ProbeSet, ProbeSetXRef)
left join ProbeSetSE on
(ProbeSetSE.DataId = ProbeSetData.Id AND ProbeSetSE.StrainId = ProbeSetData.StrainId)
left join NStrain on
(NStrain.DataId = ProbeSetData.Id AND
NStrain.StrainId = ProbeSetData.StrainId)
WHERE
ProbeSet.Name = '4336695' AND ProbeSetXRef.ProbeSetId = ProbeSet.Id AND
ProbeSetXRef.ProbeSetFreezeId = ProbeSetFreeze.Id AND
ProbeSetFreeze.Name = 'UMUTAffyExon_0209_RMA' AND
ProbeSetXRef.DataId = ProbeSetData.Id AND
ProbeSetData.StrainId = Strain.Id
Order BY
Strain.Name;
Prototocol from
Every table update has to follow the template:
mysql 1.5K Jun 27 2021 ProbeSetData.frm mysql 63G Dec 4 22:15 ProbeSetData.MYD mysql 118G Dec 28 08:52 ProbeSetData.MYI mysql 1.5K Jan 1 20:45 mytest.frm mysql 20G Jan 1 21:34 mytest.ibd
If you have enough space, you can create a copy of the actual table and do the work on that:
Original:
| ProbeSetData | CREATE TABLE 'ProbeSetData' (
'Id' int(10) unsigned NOT NULL DEFAULT 0,
'StrainId' int(20) NOT NULL,
'value' float NOT NULL,
UNIQUE KEY 'DataId' ('Id','StrainId'),
KEY 'strainid' ('StrainId')
) ENGINE=MyISAM DEFAULT CHARSET=latin1 |
New:
DROP table mytest2; CREATE table mytest2( Id INT unsigned NOT NULL DEFAULT 0, StrainId MEDIUMINT unsigned NOT NULL DEFAULT 0, value float NOT NULL, primary key(Id,StrainId), key (StrainId)) CHARACTER SET utf8mb4 ;
Note loading latin1 is slower than utf8! I tried.
And fill it killing quite a few birds with one stone :). You may want to disable the binary logger before running:
INSERT INTO mytest2 (Id,StrainId,value) SELECT Id,StrainId,value FROM ProbeSetData LIMIT 51631478;
Makes a table at 1% of size - a table of 2.3GB. I wonder why the full thing took up 330Gb before we ran out of disk space.
Show progress
SELECT TABLE_SCHEMA,TABLE_NAME,DATA_LENGTH,AVG_ROW_LENGTH,TABLE_ROWS FROM information_schema.TABLE S WHERE DATA_LENGTH>1000000000 order by data_length;
May add order by Id,StrainID next time.
Once the process is done, you can rename the tables:
DROP ProbeSetData; RENAME TABLE tbl_name TO new_tbl_name, tbl_name2 TO ProbeSetData;
SHOW CREATE TABLE ProbeSetData;
This is a really simple table
| ProbeSetData | CREATE TABLE 'ProbeSetData' (
'Id' int(10) unsigned NOT NULL DEFAULT 0,
'StrainId' int(20) NOT NULL,
'value' float NOT NULL,
UNIQUE KEY 'DataId' ('Id','StrainId'),
KEY 'strainid' ('StrainId')
) ENGINE=MyISAM DEFAULT CHARSET=latin1 |
MariaDB [db_webqtl]> select * from ProbeSetData limit 2; +----+----------+-------+ | Id | StrainId | value | +----+----------+-------+ | 1 | 1 | 5.742 | | 1 | 2 | 5.006 | +----+----------+-------+ 2 rows in set (0.000 sec)
But large
MariaDB [db_webqtl]> select count(*) from ProbeSetData; +------------+ | count(*) | +------------+ | 5141631478 | +------------+ 1 row in set (0.000 sec)
This parses the full table (MYISAM version)
MariaDB [db_webqtl]> select count(distinct id) from ProbeSetData; +--------------------+ | count(distinct id) | +--------------------+ | 48119047 | +--------------------+ 1 row in set (4 min 40.035 sec)
Do this also on Penguin2.
| ProbeSetData | CREATE TABLE 'ProbeSetData' (
'Id' int(10) unsigned NOT NULL DEFAULT 0,
'StrainId' smallint(5) unsigned NOT NULL DEFAULT 0,
'value' float NOT NULL,
UNIQUE KEY 'DataId' ('Id','StrainId')
) ENGINE=InnoDB DEFAULT CHARSET=latin1 |
The table layout is a bit puzzling.
MariaDB [db_webqtl]> select max(StrainId) from ProbeSetData; +---------------+ | max(StrainId) | +---------------+ | 65777 | +---------------+ MariaDB [db_webqtl]> select max(Id) from ProbeSetData; +----------+ | max(Id) | +----------+ | 91406791 | +----------+
For StrainID MediumInt is a better fit with unsigned range is 0 to 16777215 and Id should be a INT (instead of BIGINT for the time being). That will make the table+indices smaller and faster. Note that MediumInt does not save RAM, just disk space. But that can impact IO performance by 25%.
When a value overflows MariaDB will throw an error. Therefore it is safe to pick the smallest possible size. So even if we grow out of values we can quickly resize the table in, say, 10 years. That is, if we are still using SQL.
rg ProbeSetData --color=always |less -R
rg ProbeSetData --type=py -l|fzf --preview="rg --color=always -A 20 ProbeSetData {}" --preview-window=right:85%:wrap
No surprises from the code scan.
Some select statement and maybe a page of GN2.
The tests we can share from ProbeSetXRef.
Page of GN2 is the mapping page, e.g.
ALTER TABLE ProbeSetData ADD PRIMARY KEY(Id,StrainId);
(no quotes for column names)
SHOW INDEX FROM ProbeSetData; DROP INDEX ProbeSetId ON ProbeSetData;
ALTER TABLE ProbeSetData ENGINE = InnoDB;
ALTER TABLE ProbeSetSE CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci;
Use utf8 instead!
For those cases see bottom of move-to-innodb.gmi.
OPTIMIZE NO_WRITE_TO_BINLOG TABLE ProbeSetData;
After running the optimiser rename the tables
RENAME TABLE orig TO orig_old, mytest TO orig;
ProbeSetData is the final table that needs to be migrated to innodb. The main problem is that it is huge and transforming the table runs out of disk space. In the first step I made a borg snapshot on
root@tux01:/export2/backup# borg create --progress --stats borg-mariadb-snapshot::migrate1a /export/local/home/mariadb/
next copy the database to a new partition:
root@tux01:/export4/local/home/mariadb/database/db_webqtl# rsync -vaP /var/lib/mysql/db_webqtl/* . --delete --bwlimit=20M
Note I throttle the speed because the system can become quite unusable at full copy speed. Next I stopped Mariadb and made sure the copy is completed. After restarding mariadb I could continue work on the copy using a guix shell instance as described in
Steps were as a normal user
tux01:/export4/local/home/mariadb$ ~/opt/guix-pull/bin/guix pull -p ~/opt/guix-latest . ~/opt/guix-latest/etc/profile mkdir var guix shell -C -N coreutils sed mariadb --share=var=/var mysqld_safe --datadir='./database' --user=$USER --nowatch --socket=/var/run/mysqld/mysqld.sock mysql --socket=/var/run/mysqld/mysqld.sock -uwebqtlout -p db_webqtl
OK, now it is running and we can start experimenting with the table outside the main database setup. Remember we had
ProbeSetData | CREATE TABLE 'ProbeSetData' (
'Id' int(10) unsigned NOT NULL DEFAULT 0,
'StrainId' int(20) NOT NULL,
'value' float NOT NULL,
UNIQUE KEY 'DataId' ('Id','StrainId'),
KEY 'strainid' ('StrainId')
) ENGINE=MyISAM DEFAULT CHARSET=latin1
DROP INDEX strainid ON ProbeSetData; DROP INDEX DataId ON ProbeSetData;
of course it starts making a copy of the whole table and takes hours(!) This is why we need over 200Gb free both on the DB directory and the tempdir of the mariadb server.
select count(Id) from ProbeSetData; +------------+ | count(Id) | +------------+ | 5173425135 | +------------+ MariaDB [db_webqtl]> select max(Id),max(StrainId) from ProbeSetData; +----------+---------------+ | max(Id) | max(StrainId) | +----------+---------------+ | 92199993 | 71224 | +----------+---------------+ MariaDB [db_webqtl]> select * from ProbeSetData limit 4; +----+----------+-------+ | Id | StrainId | value | +----+----------+-------+ | 1 | 1 | 5.742 | | 1 | 2 | 5.006 | | 1 | 3 | 6.079 | | 1 | 4 | 6.414 | +----+----------+-------+
ALTER TABLE ProbeSetData MODIFY StrainId mediumint UNSIGNED NOT NULL;
Now the table is 58Gb without indices. Convert to innodb and add indices
CREATE INDEX id_index ON ProbeSetData(Id); CREATE INDEX strainid_index ON ProbeSetData(StrainId); ALTER TABLE ProbeSetData ENGINE = InnoDB;
Now we are at 150Gb.
The conversion to InnoDB, annoyingly, leads to an exceedingly large file. I think we'll keep the myisam version for ProbeSetData for the time being. It does not do text comparison, so we should see no slow downs mixing with InnoDB. From this exercise I know we can reduce the size a little. Also ProbeSet can probably be made smaller by combining id+strainid where id is dataset id and perhaps even reducing the value size by compressing it into a smaller float. Using lmdb may not be that beneficial because, even though the data is a vector, it will be costly to traverse by strainid. Checking the code base that does not happen much without selecting on Id first. So keys on Id may work for vectors of strain id and values. Introducing lmdb would require rewriting about 5 queries in GN1 and GN2 each.
I found it is a bad idea to remove large .ibd files by hand because mariadb wants to recreate them to play the transaction log.