Reference architecture for a write-intensive MySQL deployment

We designed Percona Cloud Tools (both hardware and software setup) to handle a very high-intensive MySQL write workload. For example, we already observe inserts of 1bln+ datapoints per day. So I wanted to share what kind of hardware we use to achieve this result.

Let me describe what we use, and later I will explain why.


  • Chassis: Supermicro SC825TQ-R740LPB 2U Rackmount Chassis
  • Motherboard: Supermicro X9DRI-F dual socket
  • CPU: Dual Intel Xeon Ivy Bridge E5-2643v2 (6x 3.5Ghz cores, 12x HT cores, 25M L3)
  • Memory: 256GB (16x 16GB 256-bit quad-channel) ECC registered DDR3-1600
  • Raid: LSI MegaRAID 9260-4i 4-port 6G/s hardware RAID controller, 512M buffer
  • MainStorage: PCIe SSD HGST FlashMAX II 4.8TB
  • Secondary Storage (OS, logs): RAID 1 over 2x 3TB hard drives


When selecting hardware for your application, you need to look at many aspects – typically you’re looking for a solution for which you already have experience in working with and has also proved to be the most efficient option. For us it has been as follows:

Cloud vs Bare Metal
We have experience having hardware hosted at the data center as well as cash for upfront investments in hardware so we decided to go for physical self-hosted hardware instead of the cloud. Going this route also gave us maximum flexibility in choosing a hardware setup that was the most optimal for our application rather than selecting one of the stock options.

Scale Up vs Scale Out
We have designed a system from scratch to be able to utilize multiple servers through sharding – so our main concern is choosing the most optimal configuration for the server and provisioning servers as needed. In addition to raw performance we also need to consider power usage and overhead of managing many servers which typically makes having slightly more high-end hardware worth it.

Resource Usage
Every application uses resources in different ways so an optimal configuration will be different depending on your application. Yet all applications use the same resources you need to consider. Typically you want to plan for all of your resources to be substantially used – providing some margin for spikes and maintenance.


  • Our application processes a lot of data and uses the TokuDB storage engine which uses a lot of CPU for compression, so we needed powerful CPUs.
  • Many MySQL functions are not parallel, think executing single query or Alter table so we’re going for CPU with faster cores rather than larger amount of cores. The resulting configuration with 2 sockets giving 12 cores and 24 threads is good enough for our workloads.
  • Lower end CPUs such as Xeon E3 have very attractive price/performance but only support 32GB of memory which was not enough for our application.


  • For database boxes memory is mainly used as a cache, so depending on your application you may be better off investing in memory or storage for optimal performance. Check out this blog post for more details.
  • Accessing data in memory is much faster than even on the fastest flash storage so it is still important.
    For our workload having recent data in memory is very important so we get as much “cheap” memory as we can populating all 16 slots with 16GB dimms which have attractive cost per GB at this point.

There are multiple uses for the storage so there are many variables to consider

  • Bandwidth
    • We need to be able access data on the storage device quickly and with stable response time. HGST FlashMax II has been able to meet these very demanding needs.
  • Endurance
    • When using flash storage you need to worry about endurance – how much beating with writes flash storage can handle before it wears out. Some low cost MLC SSDs would wear out in the time frame of weeks if being written with maximum speed. HGST FlashMax II has endurance rating of 10 Petabytes written (for a random workload) – 30 Petabytes written (for a sequential workload)
    • We also use TokuDB storage engine which significantly reduces amount of writes compared to Innodb.
  • Durability
    • Does the storage provide true durability with data guaranteed to be persisted when write is acknowledged at the operating system level when power goes down or is loss possible?
      We do not want to risk database corruption in case of power failure so we were looking for storage solution which guarantees durability.
      HGST FlashMax II guarantees durability which has been confirmed by our stress tests.
  • Size
    • To scale application storage demands you need to scale both number of IO operations storage can handle and storage size. For flash storage it is often the size which becomes limiting factor.
      HGST FlashMax II 4.8 TB capacity is best available on the market which allows us to go “All Flash” and achieve very quick data access to all our data set.
  • Secondary Storage
    • Not every application need requires flash storage properties.
    • We have secondary storage with conventional drives for operating system and logs.
      Sequential read/write pattern works well with low cost conventional drives and also allow us to increase flash life time, having it handling less writes.
    • We’re using RAID with BBU for secondary storage to be able to have fully durable binary logs without paying high performance penalty.

There are arguments that SATA SSD provides just a good enough performance for MySQL and there is no need for PCIe. While these arguments are valid in one dimension, there are several more to consider.

First, like I said PCIe SSD still provides a best absolute response time and it is an important factor for an end user experience in SaaS systems like Percona Cloud Tools.
Second, consider maintenance operations like backup, ALTER TABLES or slave setups. While these operations are boring and do not get as much attention as a response time or throughput in benchmarks, it is still operations that DBAs performs basically daily, and it is very important to finish a backup or ALTER TABLE in a predictable time, especially on 3-4TB datasize range. And this is where PCIe SSD performs much better than SATA SSDs. For SATA SSD, especially bigger size, write endurance is another point of concern.

Why TokuDB engine?
The TokuDB engine is the best when it comes to insert operations to a huge dataset, and few more factors makes it a no-brainer:

  • TokuDB compression is a huge win. I estimate into this storage ( FlashMAX II 4.8TB) we will fit about 20-30TB of raw data.
  • TokuDB is SSD friendly, as it performs much less data writes per INSERT operation than InnoDB, which greatly extends SSD (which is, well, expensive to say the least) lifetime.

The post Reference architecture for a write-intensive MySQL deployment appeared first on MySQL Performance Blog.

Read more at: //

Tools and tips for analysis of MySQL’s Slow Query Log

MySQL's Slow Query LogMySQL has a nice feature, slow query log, which allows you to log all queries that exceed a predefined about of time to execute. Peter Zaitsev first wrote about this back in 2006 – there have been a few other posts here on the MySQL Performance Blog since then (check this and this, too) but I wanted to revisit his original subject in today’s post.

Query optimization is essential for good database server performance and usually DBAs need to ensure the top performance possible for all queries. In MySQL, the desirable way is to generate a query log for all running queries within a specific time period and then run a query analysis tool to identify the bad queries. Percona Toolkit’s pt-query-digest is one of the most powerful tools for SQL analysis. That’s because pt-query-digest can generate a very comprehensive report that spots problematic queries very efficiently. It works equally well with Oracle MySQL server. This post will focus mainly on pt-query-digest.

Slow query log is great at spotting really slow queries that are good candidates for optimization. Beginning with MySQL 5.1.21, the minimum value is 0 for long_query_time, and the value can be specified to a resolution of microseconds. In Percona Server additional statistics may be output to the slow query log. You can find the full details here. For our clients, we often need to identify queries that impact an application the most. It does not always have to be the slowest queries – queries that runs more frequently with lower execution time per call put more load on a server than queries running with lower frequency. We of course want to get rid of really slow queries but to really optimize application throughput, we also need to investigate queries that generate most of the load. Further, if you enable option log_queries_not_using_indexes  then MySQL will log queries doing full table scans which doesn’t always reflect that the query is slow, because in some situations the query optimizer chooses full table scan rather than using any available index or probably showing all records from a small table.

Our usual recommendation is to generate the slow log with long_query_time=0. This will record all the traffic but this will be I/O intensive and will eat up disk space very quickly depending on your workload. So beware of running with long_query_time=0 for only a specific period of time and revert it back to logging only very slow queries. In Percona Server there is nice option where you can limit the rate of logging, log_slow_rate_limit is the option to handle it. Filtering slow query log is very helpful too in some cases e.g. if we know the main performance issue is table scans we can log queries only doing full table scans or if we see I/O is bottleneck we can collect queries doing full scans and queries creating on disk temporary tables. Again, this is only possible in Percona Server with the log_slow_filter option. Also, you may want to collect everything on slow query log and then filter with pt-query-digest. Depending on I/O capacity, you might prefer one or another way, as collecting everything in slow query log allows us to investigate other queries too if needed. Finally, use pt-query-digest to generate an aggregate report over slow query log which highlights the problematic part very efficiently. Again, pt-query-digest can bring up server load high so our usual recommendation on it is to move slow query log to some staging/dev server and run pt-query-digest over there to generate the report.

Note: changing the long_query_time parameter value only affects newly created connections to log queries exceeds long_query_time threshold. In Percona Server there is feature which changes variable scope to global instead of local. Enabling slow_query_log_use_global_control  log queries for connected sessions too after changing long_query_time parameter threshold. You can read more about this patch here.

I am not going to show you a detailed report of pt-query-digest and explain each part of it here, because it is well defined already by my colleague Ovais Tariq in this post. However, I will show you some of the other aspects of pt-query-digest tool here.

Let me show you code snippets that enable slow query log for only a specific time period with long_query_time=0 and log_slow_verbosity to ‘full’. log_slow_verbosity is a Percona Server variable which logs extra stats such as information on query cache, Filesort, temporary tables, InnoDB statistics etc. Once you are done collecting logs, revert back the values for long_query_time to the previous value, and finally run pt-query-digest on the log to generate report. Note: run the below code in same MySQL session.

-- Save previous settings
mysql> SELECT @@global.log_slow_verbosity INTO @__log_slow_verbosity;
mysql> SELECT @@global.long_query_time INTO @__long_query_time;
mysql> SELECT @@global.slow_query_log INTO @__slow_query_log;
mysql> SELECT @@global.log_slow_slave_statements INTO @__log_slow_slave_statements;
-- Keep this in safe place, we'll need to run pt-query-digest
mysql> SELECT NOW() AS "Time Since";
-- Set values to enable query collection
mysql> SET GLOBAL slow_query_log_use_global_control='log_slow_verbosity,long_query_time';
mysql> SET GLOBAL log_slow_verbosity='full';
mysql> SET GLOBAL slow_query_log=1;
mysql> SET GLOBAL long_query_time=0;
mysql> SET GLOBAL log_slow_slave_statements=1;
-- Verify settings are OK
mysql> SELECT @@global.long_query_time, @@global.slow_query_log, @@global.log_slow_verbosity;
-- wait for 30 - 60 minutes
-- Keep this one too, also for pt-query-digest
mysql> SELECT NOW() AS "Time Until";
-- Revert to previous values
mysql> SET GLOBAL [email protected]__slow_query_log;
mysql> SET GLOBAL [email protected]__long_query_time;
mysql> SET GLOBAL [email protected]__log_slow_verbosity; -- if percona server
mysql> SET GLOBAL [email protected]__log_slow_slave_statements;
-- Verify settings are back to previous values
mysql> SELECT @@global.long_query_time, @@global.slow_query_log, @@global.log_slow_verbosity, @@global.slow_query_log_file;
-- Then with pt-query-digest run like (replace values for time-since, time-until and log name)
$ pt-query-digest --since='<time-since>' --until='<time-until>' --limit=100% /path/to/slow_query_log_file.log > /path/to/report.out
-- If you're not using Percona Server then you need to remove all references to log_slow_verbosity, slow_query_log_use_global_control and log_slow_slave_statements (priot MySQL 5.6).

My colleague Bill Karwin wrote bash script that does almost the same as the above code. You can find the script to collect slow logs here. This script doesn’t hold connection to the database session while you wait for logs to accumulate and it sets all the variables back to the state they were before. For full documentation view this.

Further, you can also get explain output into the report from the pt-query-digest tool. For that you need to use –explain parameter similar to as follows.

$ pt-query-digest --explain u=<user>,p=<password>,h=<hostname> /path/to/slow.log > /path/to/report.out

Explain output in query report will get you all the information for query execution plan and explain output signal towards how that particular query going to be executed. Note that, if you execute pt-query-digest over slow query log other than originated server of slow query log as I mentioned above e.g. staging/dev you may get different execution path for the query in the report or lower number of rows to examined, etc., because usually staging/dev servers has different data distribution, different MySQL versions, or different indexes. MySQL explain adds overhead as queries needs to be prepared on the server to generate intended query execution path. For this reason, you may want to run pt-query-digest with –explain on a production replica.

It’s worth mentioning that logging queries with log_slow_verbosity in Percona Server is really handy as it shows lots of additional statistics and it is more helpful in situations when the explain plan reports a different execution path than when the query is executed. On that particular topic, you may want to check this nice post.

pt-query-digest also supports filters. You can read more about it here. Let me show you an example. The following command will discard everything apart from insert/update/delete queries in pt-query-digest output report.

$ pt-query-digest --filter '$event->{arg} =~ m/^(insert|update|delete)/i' --since='<time-since>' --until='<time-until>' --limit=100% /path/to/slow_query_log_file.log > /path/to/report.out

If you’re looking for some GUI tools for pt-query-digest then I would recommend reading this nice blogpost from my colleague Roman. Further, our CEO Peter Zaitsev also wrote a post recently where he shows the comparison between performance_schema and slow query log. Check here for details.

In related new, Percona recently announced Percona Cloud Tools, the next generation of tools for MySQL. It runs a client-side agent (pt-agent) which runs pt-query-digest on the server with some intervals and uploads the aggregated data to the Percona Cloud Tools API which process it further.  Query Analytics is one tool from the Percona Cloud Tools that provides advanced query metrics. It  is a nice visualization tool. You may be interested to learn more about it here, and it’s also worth viewing this related webinar about Percona Cloud Tools from our CTO Vadim Tkachenko.

pt-query-digest from Percona Toolkit is a versatile (and free) tool for slow query log analysis. It provides good insight about every individual query, especially in Percona Server with log_slow_verbosity enabled, e.g. log queries with microsecond precision, log information about the query’s execution plan. On top of that, Percona Cloud Tools includes Query Analytics which provides you with good visuals about query performance and also provides a view of historical data.

The post Tools and tips for analysis of MySQL’s Slow Query Log appeared first on MySQL Performance Blog.

Read more at: //

Percona Server with TokuDB: Packing 15TB into local SSDs

Two weeks ago we released an Alpha release of Percona Server with TokuDB. Right now I am on a final stage of evaluation of TokuDB for using in our project Percona Cloud Tools and it looks promising.

What is the most attractive in TokuDB? For me it is compression, but not just compression: TokuDB provides great performance over compressed data.

In my synthetic tests I saw a compression ratio of 10:1 (TokuDB LZMA to InnoDB uncompressed), in the real production data it is less, 6:1, but still impressive.

In our servers we have 4 x SSD Crucial M500 960GB combined in RAID5, which give 2877.0 GB of usable space. With TokuDB we should be able to pack around 15TB of raw data. Of course we can try InnoDB compression, but the best we can get is 2x compression without sacrificing performance.

And of course TokuDB is transaction, fully ACID-compliant with automatic crash-recovery storage engine.

This all makes TokuDB a very attractive choice for handling terabytes of data (or as it popular to say nowadays “BigData”).

One of first operational questions we have is how to handle backups.
For backups we use LVM partitions and the mylvmbackup tool. Unfortunately Percona XtraBackup is not able to handle TokuDB tables (and probably won’t be able anytime soon). The other choice is to use TokuDB Hot back-up, available by Tokutek Enterprise Subscription. I did not test it myself, so I can’t provide any feedback.

And of course there are things which I do not fully like in TokuDB:

  • No Foreign Keys support. It is not a big issue for us, but I know for some users this is a showstopper.
  • Time-based checkpoints. You may not notice a direct effect from this, but we clearly see it in our benchmarks. Every 60 sec (default timeperiod between checkpoints) we see a drop in throughput during write-intensive benchmarks. It is very similar to drops in InnoDB we tried to solve (and are still trying), for example see Adaptive flushing in MySQL 5.6. My advice to the Tokutek team would be to also look into a fuzzy check-pointing, instead of time-based.
  • All TokuDB files are stored in a single directory, sometime with mangled filenames. This especially becomes bad in sharding or multi-tenant environments when tens of thousands of files are in the same directory

Well, I guess for now, we will take these limitations as TokuDB specific and will deal with them.

Next week we plan on a Beta release of Percona Server with TokuDB, so stay tuned!

The post Percona Server with TokuDB: Packing 15TB into local SSDs appeared first on MySQL Performance Blog.

Read more at: //