#DBHangOps 02/19/15 -- Long Query Time, Operational TokuDB, and more!
Join in #DBHangOps this Thursday, February, 19, 2015 at 11:00am pacific (19:00 GMT), to participate in the discussion about:
- Learnings from operating TokuDB
- What's a good
- Testing your backups
- MySQL 5.7 defaults suggestions
You can check out the event page at https://plus.google.com/events/cohut2qncrbkrrmbs868kjorvbo on Thursday to participate.
As always, you can still watch the #DBHangOps twitter search, the @DBHangOps twitter feed, or this blog post to get a link for the google hangout on Thursday!
See all of you on Thursday!
You can catch a livestream at:
Learnings from operating TokuDB
- Jeremy Tinley from Etsy setup (and blogged about) setting up TokuDB
- Had a mysql store that was getting bulk-loaded periodically
- Previously using Percona server
- Wanted to try out TokuDB for faster bulk load and high compression rates. As part of testing:
- Workload was predominantly insert-bound
- 190GB InnoDB tablespace was ~30GB in TokuDB
- Got a 30% speed improvement
- Setup Chef recipes for Percona server 5.6.17+ with TokuDB
- Using JMAlloc
- Disabling Transparent Huge Pages.
- Setup MySQL to not start on boot so Chef can properly disable THP
- Newer version of percona server have a script to check for all the components needed to start the server with preferred plugins (e.g. JMalloc)
- TokuDB Config defaults are very sane
- One gotcha: defaulting 80% of memory to InnoDB. TokuDB defaults to 50% for the engine and the other 50% is left for OS file cache.
- Accidentally overallocated here. Changed the InnoDB heap to 5% so that memory wasn't over-committed
- TokuDB data directory
- A lot of files with a ".tokudb" extensions
- Mostly the same monitoring:
- Is MySQL up?
- How many threads in the server?
- Most things are handled at the MySQL, not the engine level, so it works pretty easily
- Need to add some graphing for "SHOW ENGINE TOKUDB STATUS" (which is very nicely formatted and easy to parse).
- This output is moslty key=value pairs and are prefixed with common names to be easy to search
- Still a challenge figuring out which new status counters to watch for the engine and to identify contentious points
- Not an easy guide to show correlations between innodb counters and tokudb counters
- Xtrabackup doesn't work for TokuDB! :(
- Using MyLVMBackup to do basic LVM snapshots on the dataset
- Needed to patch this to do an SSH copy to a backup server on the fly instead of rsync'ing after the backup
- Overall results
- Storing a feed of all modifications to production sharded environment for being inserted into Solr for searching
- 190GB InnoDB table went to ~30GB TokuDB table
- Slightly higher CPU usage due to compression, but great for I/O bound systems
- 6x increase in backfill process
- Successful win for Etsy.
- Want to get more operational experience with MySQL 5.6 and TokuDB
- Other Questions
- Is there a hot backup tool for Tokutek?
- There's an enterprise solution that comes with a hot backup tool
- Works like a snapshot method
- What's the typical workload with respect to concurrency and locking?
- ~80% insert volume and the rest are selects for inserts into Solr
- How do you trace transaction locking and deadlocks as compared to InnoDB (e.g. ENGINE INNODB STATUS and
- There's tables in
information_schema for accessing locks
- More data on Tokutek's documentation for identifying this information
- TokuDB's sweet spots are insert speed (like bulk inserts) and compression ratios. Any noticeable weak spots that came up?
- Are point selects and range queries less efficient compared to innodb?
- Jeremy hasn't observed any of this yet, but there's definitely more learning that needs to happen
- When Shlomi did some testing some locking behavior got heavy
- Schema Changes?
- Were able to apply schema changes on the fly against the engine. Except for the thread initiating the query, everything seemed to move along well.
- When you add a column, you can begin using it, even before the alter finishes
- First time playing with MySQL 5.6
- How do you set it?
- 1 second
- It's a reasonable value that can still capture enough of things
- If you go above 1 you can still miss a whole bunch of slow queries
- 500ms these days is starting to be considered "slow"
- 0 seconds?
- You get all queries when set to this threshold
- This can give you a global view of everything going through a server
- Building on top of
performance_schema data is probably more valuable to get at this data now
- Any tools that leverage
performance_schema can help find problematic queries that may not run a long time (e.g. MySQL Enterprise Monitor)
- Instead of using the slow query log to get this data, measure traffic coming over tcpdump and evaluate from there
- VividCortex does this to try and identify per-query basis measuring of traffic
- You're not required to turn on
performance_schema or the slow query log
- MySQL 5.7 is proposing to lower this value because the default is too high (10s). Planning to drive this down to 2s
- MongoDB's equivalent to long query time is 100ms.
- Is there value in having a long transaction log?
- Not yet...but you can get at this information by probing the
- Perhaps having a
transaction_timeout configuration in a future
- In MySQL 5.7, some transaction instruments will be enabled by default in performance schema.
- You can get information about the type of transaction, isolation level, length, user, etc.
- Also planning to turn on statement history instruments so you can see the last couple statements from a transaction/connection as well