maria-discuss team mailing list archive

[Stephane VAROQUI <stephane@xxxxxxxxxxx>]

Hi,

Justin points are corrects but TokuDB perform better for OLAP if you set all your indexed to clustered.  
We have many satisfy users on TokuDB and  it's a mature storage. 

POC  well 

Stéphane Varoqui, Senior Consultant
Phone: +33 695-926-401, skype: svaroqui
http://www.mariadb.com

Le Jul 5, 2015 à 4:12 AM, Justin Swanhart a écrit :

> Hi,
> 
> TokuDB works best when the dataset is too large for memory and reads and writes are small (ie, it doesn't excel at OLAP, but is good for OLTP).  If all data fits in memory, then it performs up to 3x worse than InnoDB.  Both have row level locking, though I don't know how TokuDB takes locks if you have to scan a table.  For MVCC repeatable-read I assume it has to behave like InnoDB repeatable-read, but I don't know for sure.  You could ask about TokuDB on the Percona forums.
> 
> If you are read head, why do you need select for update?  At least do lock in share mode, then when you write, you will upgrade to X lock, with small chance of deadlock, but you will get better concurrency, significantly better.
> 
> --Justin
> 
> Regards,
> 
> --Justin
> 
> On Sat, Jul 4, 2015 at 5:36 PM, Roberto Spadim <roberto@xxxxxxxxxxxxx> wrote:
> i think the problem is select for update (i tryed with myisam/aria but
> they was using lock tables to 'solve' problems, but it didn't worked,
> obvious too much small updates/deleted and big reads is a problem with
> myisam/aria), i will check again what to do
> 
> i didn't tested tokudb yet, any experience is wellcome here, does it
> 'works' like innodb with many small writes and big reads? (i will
> check it with some days, i will try to log all queries and execute the
> same workload at another server - my laptop and check what happens)
> 
> 2015-07-03 20:17 GMT-03:00 Justin Swanhart <greenlion@xxxxxxxxx>:
> > Hi,
> >
> > You can't create a slave of the master itself, because you can't replicate
> > from/to same server_id.
> >
> > Before you consider other solutions, you should look into why InnoDB is
> > locking.  InnoDB locks are held during INSERT/UPDATE/DELETE operations and
> > during SELECT .. FOR UPDATE or SELECT .. LOCK IN SHARE MODE.  If you are
> > getting locking issues it is because multiple writers are writing to the
> > same rows, or you are using SELECT with a locking option.
> >
> > If you are doing updates or deletes, make sure the table is well indexed for
> > the operation.  InnoDB will lock all rows it traverses, so if an
> > update/delete needs a FTS, whole table is locked.  You can somewhat mitigate
> > this problem with READ COMMITTED, as it will release the locks that are not
> > needed after the scan finishes, but you will still lock many (or all) rows
> > during the scan.
> >
> > FlexCDC has an SBR replication mechanism in trunk/ (it is in
> > consumer/replication I think).  I haven't tested it in awhile but SBR
> > doesn't really change.  I can't remember if it supports filtering only a
> > specific table though.  I wrote it four or five years ago so my memory is
> > fuzzy.  I can trivially add that though if it missing.
> >
> > Or just use a Flexviews materialized view:
> >
> > call flexviews.create_mvlog('schema', 'A');
> > call flexviews.create('schema','B', 'INCREMENTAL');
> > set @mvid := LAST_INSERT_ID();
> > call flexviews.add_expr(@mvid,'COLUMN','*','all_columns');
> > call flexviews.add_table(@mvid, 'schema', 'A', 'A',NULL);
> > call flexviews.enable(@mvid);
> >
> >
> > Or use mysqlbinlog + php/python/perl to extract statements for table A, and
> > rewrite the statements to table B, and apply them.  Use SBR for this,
> > because otherwise the complexity is same as FlexCDC and you should just use
> > materialized view.
> >
> > --Justin
> >
> >
> >
> > On Fri, Jul 3, 2015 at 3:23 PM, Stephane VAROQUI <stephane@xxxxxxxxxxx>
> > wrote:
> >>
> >> Flexview can probably help with  tables  denormalization that's mainly use
> >> to avoid joins on disks
> >>
> >> About writes
> >>
> >>
> >> To keep write performance you need to have table size of all inserted
> >> tables in memory ? so usage flexview for that help only if you have table A
> >> insert heavy purged but at the same time table B would need to have less
> >> indexes  and will still get same memory issues  and instead of doing single
> >> write you would need 2 writes A and B ?
> >> Partitioning have been created for this purpose so if you can find a
> >> partition key that will always insert into the same partition than you can
> >> maintain both tables in memory .
> >>
> >> About reads .
> >>
> >> 1 - if you can put all data in memory  you will be able to get 800K reads
> >> per sec . if your count , group by , etc parse more than 1M like 32 Million
> >> it will still do 50s  to get you a result .
> >>
> >> Need a lot of slaves (cores) to satisfy 100 qps /s. that possible we have
> >> clients or users having Kilo servers to satisfy such requirements
> >>  ( maxscale can help keeping those slaves in sync when you have so many)
> >>
> >> 2- If you wan't to reduce the amount of servers needed for big data
> >> computing or you can't offer to put all data in memory and also can't wait
> >> those 32 sec you can use a column based storage, those tools are dedicated
> >> for processing multi millions data reads /s  at the price of asynchronous
> >> ETL loading
> >> You can try out InfiniDB.  documentation and binaries can be found from
> >> the MariaDB Portal
> >>
> >> 3 - Now if you think that you don't really need to read so many records
> >> like around a Million but that you can't afford to get memory , it's
> >> interesting to use TokuDB as it will parse less record per second in memory
> >> vs innodb butwill use a lot lot less disks io compare to innodb or myisam.
> >> With heavy compression and fractal tree it is a drastic io reduction.
> >>
> >> 4- Other solution to keep memory low is using a good flash storage like
> >> fusion IO , it can produce 128K reads IO per sec using1/5 in memory  that
> >> would give you still the million parse in less than few seconds.
> >>
> >> 5 - If you have access to many nodes to compute the same request , you can
> >> think spider that can split every partition into a separate server . You
> >> will end up having data back in memory and spider can help you consolidate
> >> all results performed on each server , with 32 nodes you can probably get
> >> job done on all nodes adding the time to consolidate. This would be possible
> >> on basic query plan like group and sum of a single table . but it become
> >> more interesting if you need a million record count that end up into a
> >> single partition  where data stay in the memory on the destination server .
> >>
> >> Hope it helps
> >>
> >> Stephane.
> >>
> >>
> >> Stéphane Varoqui, Senior Consultant
> >>
> >> Phone: +33 695-926-401, skype: svaroqui
> >> http://www.mariadb.com
> >>
> >> Le Jul 3, 2015 à 11:40 PM, Roberto Spadim a écrit :
> >>
> >> the main table have 29M rows (39GB, servers are "small" and old
> >> (2008), 16GB ram, 8core dual-cpu xeon, raid 10 hd with 4 sata disks
> >> 250gb each, 4 1Gbps network card )
> >> 20+ process updating, deleting, inserting (oltp, 1000+ qps, i think it
> >> will not grow with time)
> >> 5+ process running olap with big selects/group/order/min/max/sum/count
> >> etc (some queries take more than 5 minutes to end, it's not a problem
> >> to user, but a problem when use oltp+olap apps at same table)
> >>
> >> today using replication to another server or mysqld process at same
> >> machine i don't have problem, olap and oltp runs nice with different
> >> mysqld process
> >> but when i try oltp + olap at same table at same mysqld, locks begin a
> >> problem, i was thinking about creating a replication at same server
> >> but to different table (replicate table a to table b)
> >>
> >>
> >>
> >> 2015-07-03 18:28 GMT-03:00 Stephane VAROQUI <stephane@xxxxxxxxxxx>:
> >>
> >> Hi,
> >>
> >>
> >> How many writes do you have ? How much do you plan ?
> >>
> >> When you read how many records and what is your business case  ?
> >>
> >>
> >> /stephane
> >>
> >>
> >> Stéphane Varoqui, Senior Consultant
> >>
> >>
> >> Phone: +33 695-926-401, skype: svaroqui
> >>
> >> http://www.mariadb.com
> >>
> >>
> >> Le Jul 3, 2015 à 11:23 PM, Roberto Spadim a écrit :
> >>
> >>
> >> i didn't tested but spider have HA/federate, maybe i could use it?
> >>
> >>
> >> 2015-07-03 18:00 GMT-03:00 Federico Razzoli <federico_raz@xxxxxxxx>:
> >>
> >>
> >> I'm not sure. Flexviews should be a good solution.
> >>
> >>
> >>
> >> Or you can used triggers to "replicate" the table instantly.
> >>
> >>
> >>
> >> I don't understand your idea with SPIDER, how will it help you?
> >>
> >>
> >>
> >> Regards
> >>
> >>
> >> Federico
> >>
> >>
> >>
> >>
> >>
> >> --------------------------------------------
> >>
> >>
> >> Ven 3/7/15, Roberto Spadim <roberto@xxxxxxxxxxxxx> ha scritto:
> >>
> >>
> >>
> >> Oggetto: [Maria-discuss] doubt - replication at same mysqld process
> >>
> >>
> >> A: "Maria Discuss" <maria-discuss@xxxxxxxxxxxxxxxxxxx>
> >>
> >>
> >> Data: Venerdì 3 luglio 2015, 22:23
> >>
> >>
> >>
> >> hi guys, i have a doubt about
> >>
> >>
> >> replciation on same machine, i never did
> >>
> >>
> >> this before
> >>
> >>
> >>
> >>
> >> i have a table running many writes and few reads, and
> >>
> >>
> >> another process
> >>
> >>
> >> start reading a lot, my today solution is replication on two
> >>
> >>
> >> servers
> >>
> >>
> >> (on same machine or other machine), the point is... could i
> >>
> >>
> >> replicate
> >>
> >>
> >> in same server (with only one server running / only one
> >>
> >>
> >> mysqld
> >>
> >>
> >> process) ?
> >>
> >>
> >>
> >> something like change table A, and a background process
> >>
> >>
> >> replicate to
> >>
> >>
> >> table B? "many writes" will write at table A, reads will
> >>
> >>
> >> read table B
> >>
> >>
> >> (read can be out of sync)
> >>
> >>
> >>
> >> innodb is locking a lot of rows, and myisam/aria is locking
> >>
> >>
> >> table a
> >>
> >>
> >> lot, both engines i have problem with lock, i consider
> >>
> >>
> >> replication to
> >>
> >>
> >> another mysqld process as the only solution, but i'm
> >>
> >>
> >> considering
> >>
> >>
> >> running only one mysqld process (if possible)
> >>
> >>
> >> i was thinking about something like HA in spider, but i
> >>
> >>
> >> didn't tested,
> >>
> >>
> >> maybe with flexviewcdc i could have a materialized view "B"
> >>
> >>
> >> of table
> >>
> >>
> >> A?
> >>
> >>
> >>
> >> any idea/help is wellcome
> >>
> >>
> >>
> >> --
> >>
> >>
> >> Roberto Spadim
> >>
> >>
> >>
> >> _______________________________________________
> >>
> >>
> >> Mailing list: https://launchpad.net/~maria-discuss
> >>
> >>
> >> Post to     : maria-discuss@xxxxxxxxxxxxxxxxxxx
> >>
> >>
> >> Unsubscribe : https://launchpad.net/~maria-discuss
> >>
> >>
> >> More help   : https://help.launchpad.net/ListHelp
> >>
> >>
> >>
> >>
> >>
> >>
> >> --
> >>
> >> Roberto Spadim
> >>
> >> SPAEmpresarial - Software ERP
> >>
> >> Eng. Automação e Controle
> >>
> >>
> >> _______________________________________________
> >>
> >> Mailing list: https://launchpad.net/~maria-discuss
> >>
> >> Post to     : maria-discuss@xxxxxxxxxxxxxxxxxxx
> >>
> >> Unsubscribe : https://launchpad.net/~maria-discuss
> >>
> >> More help   : https://help.launchpad.net/ListHelp
> >>
> >>
> >>
> >>
> >>
> >>
> >> --
> >> Roberto Spadim
> >> SPAEmpresarial - Software ERP
> >> Eng. Automação e Controle
> >>
> >>
> >>
> >> _______________________________________________
> >> Mailing list: https://launchpad.net/~maria-discuss
> >> Post to     : maria-discuss@xxxxxxxxxxxxxxxxxxx
> >> Unsubscribe : https://launchpad.net/~maria-discuss
> >> More help   : https://help.launchpad.net/ListHelp
> >>
> >
> 
> 
> 
> --
> Roberto Spadim
> SPAEmpresarial - Software ERP
> Eng. Automação e Controle
>