We are going to move from myisam to innodb. Penguin2 has been happily running innodb for some time. Main problem are fulltext columns, the text from Trello is captured below. This is for the following tables * ProbeSet * GeneRIF_BASIC * pubmedsearch Initial good candidates are (from issues/database-not-responding): * Good candidates + 2.1G Dec 4 22:15 ProbeSetXRef.MYD + 2.3G Dec 18 14:56 ProbeSet.MYD - with fulltext column + 2.6G Aug 27 2019 ProbeSE.MYD (used?) + 7.1G Nov 2 05:07 ProbeSetSE.MYD (done) + 11G Aug 27 2019 ProbeData.MYD (used?) + 63G Dec 4 22:15 ProbeSetData.MYD I am starting with the two SE tables first - because they are small. Actually ProbeData and ProbeSE (containing the Affy data) are not referenced in GN2. I need to check that. ## Report With the SQL database we need to move from myisam to innodb format, mostly to stop the problem of full table locks. Also I expect the occasional crashes we see to go away. Today, as a start, I moved the ProbeSetSE table to innodb. The result is that the disk representation is 3x the size and a full table scan takes 3x the time (somewhat unsurprising). This may impact correlation, i.e. those routines that go through all the data. We'll have to test that carefully. For the SE I don't think we do that, so it is no biggie. There are no real performance gains that I can tell, though for the mapping page I see no slowing down either. I'll need to free up space on the NVME storage to do the larger tables. I also took the opportunity to add a primary key because InnoDB requires it and I changed the *table* language from latin1-sweden to utf8. These are major upgrades. There are no real performance gains that I can tell, though for the mapping page I see no slowing down either. I'll need to free up space on the NVME storage to do the larger tables. I also took the opportunity to add a primary key because InnoDB requires it and I changed the *table* language from latin1-sweden to utf8. These are major upgrades. In the coming week I'll have to reboot the server to add a disk, move stuff across to free up space on the NVME, and convert a few more tables. I aim to get the largest tables done - because these are the one that hurt when locked: * Good candidates with update times + 2.1G Dec 4 22:15 ProbeSetXRef.MYD + 2.3G Dec 18 14:56 ProbeSet.MYD - with fulltext column + 2.6G Aug 27 2019 ProbeSE.MYD + 7.1G Nov 2 05:07 ProbeSetSE.MYD + 11G Aug 27 2019 ProbeData.MYD + 63G Dec 4 22:15 ProbeSetData.MYD (note each is about double that size because of indexes) ProbeSetSE is done. So ProbeData and ProbeSetData are the main candidates right now. Note I set buffers to 16GB for now ``` +innodb_buffer_pool_size=16G +innodb_ft_min_token_size=3 +# innodb_use_sys_malloc=0 +innodb_file_per_table=ON ``` ## Check database ``` time mysqlcheck -c -u webqtlout -pwebqtlout db_webqtl ``` To check one single table Probe ``` root@tux01:/var/lib/mysql/db_webqtl# mysqlcheck -c db_webqtl -u webqtlout -pwebqtlout Probe db_webqtl.Probe OK ``` Make sure we have not a FULLTEXT column we do not know about (note it needs backquotes around ls): ``` root@tux01:/var/lib/mysql/db_webqtl# for x in ls -1 *.MYD|sed -e 's,\.MYD,,' ; do echo $x ; mysql -u webqtlout -pwebqtlout db_webqtl -e "select group_concat(distinct column_name) from information_schema.STATISTICS where table_schema = 'db_webqtl' and table_name = '$x' and index_type = 'FULLTEXT'" ; done|less GeneRIF_BASIC ProbeSet pubmedsearch ``` still the same. ## Create a test Start with ProbeSetSE ``` -rw-rw---- 1 mysql mysql 8.1G Dec 28 08:39 ProbeSetSE.MYI -rw-rw---- 1 mysql mysql 7.1G Nov 2 05:07 ProbeSetSE.MYD -rw-rw---- 1 mysql mysql 8.5K Feb 3 2021 ProbeSetSE.frm ``` ``` MariaDB [db_webqtl]> select * from ProbeSetSE limit 2; +--------+----------+----------+ | DataId | StrainId | error | +--------+----------+----------+ | 1 | 1 | 0.681091 | | 1 | 2 | 0.361151 | +--------+----------+----------+ 2 rows in set (0.001 sec) MariaDB [db_webqtl]> select count(*) from ProbeSetSE limit 2; +-----------+ | count(*) | +-----------+ | 688744613 | +-----------+ 1 row in set (0.000 sec) ``` ``` MariaDB [db_webqtl]> flush tables ProbeSetSE; Query OK, 0 rows affected (0.002 sec) ``` ``` MariaDB [db_webqtl]> select count(*) from ProbeSetSE where error<0.36; +-----------+ | count(*) | +-----------+ | 601603553 | +-----------+ 1 row in set (1 min 1.189 sec) ``` Some testing shows strainid and error are not indexed. Test query on myisam: ``` MariaDB [db_webqtl]> select count(*) from ProbeSetSE where strainid<20 and error<0.10; +----------+ | count(*) | +----------+ | 61625074 | +----------+ 1 row in set (58.301 sec) ``` Check index and primary key (PK): ``` SHOW CREATE TABLE ProbeSetSE; | ProbeSetSE | CREATE TABLE "ProbeSetSE" ( "DataId" int(10) unsigned NOT NULL DEFAULT 0, "StrainId" smallint(5) unsigned NOT NULL DEFAULT 0, "error" float NOT NULL, UNIQUE KEY "DataId" ("DataId","StrainId") ) ENGINE=MyISAM DEFAULT CHARSET=latin1 | ``` Note the latin1 we don't need. Also innodb needs a primary key. The mapping page retrieves SE through a call to ``` retrieve_sample_data(this_trait, dataset) ``` we also have an API endpoint ``` @app.route("/api/v_{}/sample_data/".format(version)) @app.route("/api/v_{}/sample_data/.".format(version)) def all_sample_data(dataset_name, file_format="csv"): ``` But I can't get it to work for a trait. ``` curl "http://genenetwork.org/api/v_pre1/sample_data/1427571_at" ``` so, instead, I exported the CSV from => http://genenetwork.org/show_trait?trait_id=1427571_at&dataset=HC_M2_0606_P It took the server 2.02924s seconds to process this page. It took your browser 1.577 second(s) to render this page. This is our test 'setup'. ## Backup and convert table On Tux01 the database is hosted on a drive with 111 GB free. Not enough for all conversions. There should be an extra drive in there which requires configuration with reboot. But first we can convert this small table. Backups we have already automated. But I'll add ``` tar cvzf /home/wrk/ProbeSetSE.tgz ProbeSetSE* ``` To convert to InnoDB we should: * ascertain primary key * change charset * change engine to InnoDB * make space on disk drive(s) ### Set primary key For ProbeSetSE note the UNIQUE key was already defined. So set a primary key: ``` ALTER TABLE ProbeSetSE ADD PRIMARY KEY(DataId,StrainId); Query OK, 688744613 rows affected (15 min 13.830 sec) Records: 688744613 Duplicates: 0 Warnings: 0 ``` Unsurprisingly the index grew ``` -rw-rw---- 1 mysql mysql 16G Dec 28 11:06 ProbeSetSE.MYI ``` ### Change charset ``` MariaDB [db_webqtl]> SHOW CHARACTER SET LIKE 'utf8mb4'; +---------+---------------+--------------------+--------+ | Charset | Description | Default collation | Maxlen | +---------+---------------+--------------------+--------+ | utf8mb4 | UTF-8 Unicode | utf8mb4_general_ci | 4 | +---------+---------------+--------------------+--------+ 1 row in set (0.000 sec) ``` ``` MariaDB [db_webqtl]> ALTER TABLE ProbeSetSE CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci; Stage: 2 of 2 'Enabling keys' 0% of stage done Query OK, 688744613 rows affected (15 min 14.380 sec) Records: 688744613 Duplicates: 0 Warnings: 0 ``` For this table it has no effect since there are not text fields. Still this looks good: ``` SHOW CREATE TABLE ProbeSetSE; | ProbeSetSE | CREATE TABLE "ProbeSetSE" ( "DataId" int(10) unsigned NOT NULL DEFAULT 0, "StrainId" smallint(5) unsigned NOT NULL DEFAULT 0, "error" float NOT NULL, PRIMARY KEY ("DataId","StrainId"), UNIQUE KEY "DataId" ("DataId","StrainId") ) ENGINE=MyISAM DEFAULT CHARSET=utf8mb4 | ``` Looking good! Now there are two indexes which are the same to ``` MariaDB [db_webqtl]> DROP INDEX DataId ON ProbeSetSE; ``` ``` MariaDB [db_webqtl]> show index from ProbeSetSE; 2 rows in set (0.000 sec) ``` ### Convert to innodb So, on to converting to innodb => https://mariadb.com/kb/en/converting-tables-from-myisam-to-innodb/ => https://dataedo.com/kb/query/mysql/list-innodb-tables Show existing innodb tables Unfortunately we can't swith to 4k page tables because we have existing tables. We'll do that later some day in a controlled fashion. ``` MariaDB [db_webqtl]> ALTER TABLE ProbeSetSE ENGINE = InnoDB; Query OK, 688744613 rows affected (1 hour 51 min 36.273 sec) Records: 688744613 Duplicates: 0 Warnings: 0 ``` 2 hours for a lousy table conversion! The new file sizes are: ``` -rw-rw---- 1 mysql mysql 1.5K Dec 28 11:33 ProbeSetSE.frm -rw-rw---- 1 mysql mysql 51G Dec 28 13:25 ProbeSetSE.ibd ``` So the has tripled (including a new index) and during conversion it has both the old and the new on disk. I'll need to make space for this baby. Let's try OPTIMIZE ``` OPTIMIZE NO_WRITE_TO_BINLOG TABLE ProbeSetSE; ``` Ah, now the size is similar to myisam and loading the mapping page is slighty faster. ## Run tests again What about performance? ``` MariaDB [db_webqtl]> select count(*) from ProbeSetSE where strainid<20 and error<0.10; +----------+ | count(*) | +----------+ | 61625074 | +----------+ 1 row in set (3 min 22.958 sec) ``` Whoah. 3 times slower - which makes sense if you know the physical size of the data. Full table scans should be rare, but we need to make sure we don't slow them down that much! Again I exported the CSV from => http://genenetwork.org/show_trait?trait_id=1427571_at&dataset=HC_M2_0606_P It took the server 1.46351s seconds to process this page. It took your browser 2.119 second(s) to render this page and it shows practically the same results. ## Table template Prototocol from => ./move-to-innodb.gmi Every table update has to follow the template: ### Check table structure ``` SHOW CREATE TABLE MYTABLE; ``` ``` select * from MYTABLE limit 2; ``` ``` select count(*) from MYTABLE; ``` ### Check GN1,2,3 code for use of table ``` rg MYTABLE --color=always |less -R ``` ### Create test Some select statement and maybe a page of GN2. ### Create primary key ``` ALTER TABLE MYTABLE ADD PRIMARY KEY(col1,col2); ``` ### Create indices ### Convert to innodb ``` ALTER TABLE MYTABLE ENGINE = InnoDB; ``` ### Change charset ``` ALTER TABLE ProbeSetSE CONVERT TO CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci; ``` ### Update fulltext For those cases see bottom of move-to-innodb.gmi. ### Run optimiser ``` OPTIMIZE NO_WRITE_TO_BINLOG TABLE MYTABLE; ``` ### Check test ## ProbeSetXRef => ProbeSetXRef.gmi ## Notes captured from Trello: In an earlier track I wrote how to deal with Fulltext fields pjotrp 16 Oct 2019 at 10:47 Fulltext To list fulltext info for one table do: ``` select group_concat(distinct column_name) from information_schema.STATISTICS where table_schema = 'db_webqtl' and table_name = 'Probeset' and index_type = 'FULLTEXT'; ``` Or ``` SHOW CREATE TABLE ProbeSet; E.g. FULLTEXT KEY `SEARCH_GENE_IDX` (`Symbol`,`alias`), FULLTEXT KEY `SEARCH_FULL_IDX` (`Name`,`description`,`Symbol`,`alias`,`GenbankId`,`UniGeneId`,`Probe_Target_Description`), FULLTEXT KEY `RefSeq_FULL_IDX` (`RefSeq_TranscriptId`) ``` To see all: ``` root@tux02:/var/lib/mysql/db_webqtl# for x in ls -1 *.MYD|sed -e 's,\.MYD,,' ; do echo $x ; mysql -u webqtlout -pwebqtlout db_webqtl -e "select group_concat(distinct column_name) from information_schema.STATISTICS where table_schema = 'db_webqtl' and table_name = '$x' and index_type = 'FULLTEXT'" ; done ``` So we have to fix only ``` ProbeSet GeneRIF_BASIC pubmedsearch ``` ``` REPAIR TABLE ProbeSet QUICK; REPAIR TABLE GeneRIF_BASIC QUICK; REPAIR TABLE pubmedsearch QUICK; ``` Note that GN1 search only appears to use ProbeSet. Reindexing takes about 10 minutes on Tux02. After updating to ProbeSet to innodb the following query failed ``` SELECT distinct ProbeSet.Name as TNAME, 0 as thistable, ProbeSetXRef.Mean as TMEAN, ProbeSetXRef.LRS as TLRS, ProbeSetXRef.PVALUE as TPVALUE, ProbeSet.Chr_num as TCHR_NUM, ProbeSet.Mb as TMB, ProbeSet.Symbol as TSYMBOL, ProbeSet.name_num as TNAME_NUM FROM ProbeSetXRef, ProbeSet WHERE ((MATCH (ProbeSet.Name, ProbeSet.description, ProbeSet.symbol, alias, GenbankId, UniGeneId, Probe_Target_Description) AGAINST ('shh' IN BOOLEAN MODE)) or (MATCH (ProbeSet.symbol) AGAINST ('"Hx" "ShhNC" "9530036O11Rik" "Dsh" "Hhg1" "Hxl3" "M100081"' IN BOOLEAN MODE))) and ProbeSet.Id = ProbeSetXRef.ProbeSetId and ProbeSetXRef.ProbeSetFreezeId = 112 ORDER BY ProbeSet.symbol ASC; ``` with ``` ERROR 1191 (HY000): Can't find FULLTEXT index matching the column list ``` The shorter version works ``` SELECT distinct ProbeSet.Name as TNAME, 0 as thistable, ProbeSetXRef.Mean as TMEAN, ProbeSetXRef.LRS as TLRS, ProbeSetXRef.PVALUE as TPVALUE, ProbeSet.Chr_num as TCHR_NUM, ProbeSet.Mb as TMB, ProbeSet.Symbol as TSYMBOL, ProbeSet.name_num as TNAME_NUM FROM ProbeSetXRef, ProbeSet WHERE ((MATCH (ProbeSet.Name, ProbeSet.description, ProbeSet.symbol, alias, GenbankId, UniGeneId, Probe_Target_Description) AGAINST ('"Shh" "ShhNC" "9530036O11Rik" "Dsh" "Hhg1"' IN BOOLEAN MODE))) and ProbeSet.Id = ProbeSetXRef.ProbeSetId and ProbeSetXRef.ProbeSetFreezeId = 112; ``` when you remove any column in the MATCH statement we get this error. Which kinda makes sense I suppose. We need to add an index for the single ProbeSet.symbol match. Create it with ``` CREATE FULLTEXT INDEX ft_ProbeSet_Symbol ON ProbeSet(Symbol); ``` Now we have ``` FULLTEXT KEY `SEARCH_GENE_IDX` (`Symbol`,`alias`), FULLTEXT KEY `SEARCH_FULL_IDX` (`Name`,`description`,`Symbol`,`alias`,`GenbankId`,`UniGeneId`,`Probe_Target_Description`), FULLTEXT KEY `RefSeq_FULL_IDX` (`RefSeq_TranscriptId`), FULLTEXT KEY `ft_ProbeSet_Symbol` (`Symbol`) ``` and the query works. pjotrp 16 Oct 2019 at 09:24 (edited) Full text search is the least straightforward, see and Now for ProbeSet - one of the critical locked tables we need this. pjotrp 15 Oct 2019 at 16:38 (edited) ``` time mysqlcheck -c -u webqtlout -pwebqtlout db_webqtl db_webqtl.Docs warning : 1 client is using or hasn't closed the table properly status : OK real 16m52.567s ``` ## Convert to InnoDB The largest tables are ``` 1.6G Aug 27 2019 Probe.MYD 2.1G Aug 27 2019 LCorrRamin3.MYD 2.1G Dec 4 22:15 ProbeSetXRef.MYD 2.3G Dec 18 14:56 ProbeSet.MYD 2.6G Aug 27 2019 ProbeSE.MYD 7.1G Nov 2 05:07 ProbeSetSE.MYD 8.3G Aug 28 2019 SnpPattern.MYD 11G Aug 27 2019 ProbeData.MYD 11G May 22 2020 GenoData.MYD 11G Aug 27 2019 SnpAll.MYD 63G Dec 4 22:15 ProbeSetData.MYD ``` On Penguin2 we are already runing ProbeSetData as ``` 238G Jul 10 2020 ProbeSetData.ibd ``` which is pretty massive! It includes the index, which is 180G, so the difference is not that great. Also we should try a 4kb page size. Also make sure to enable innodb_file_per_table.