EmergencyEMERGENCY? Get 24/7 Help Now!

Capture database traffic using the Performance Schema

 | October 1, 2015 |  Posted In: Insight for DBAs, MySQL, Performance Schema

PREVIOUS POST
NEXT POST

Capturing data is a critical part of performing a query analysis, or even just to have an idea of what’s going on inside the database.

There are several known ways to achieve this. For example:

  • Enable the General Log
  • Use the Slow Log with long_query_time = 0
  • Capture packets that go to MySQL from the network stream using TCPDUMP 
  • Use the pt-query-digest with the –processlist parameter

However, these methods can add significant overhead and might even have negative performance consequences, such as:

Now, sometimes you just need to sneak a peek at the traffic. Nothing fancy. In that case, probably the faster and easiest way to gather some traffic data is to use pt-query-digest with the –processlist. It doesn’t require any change in the server’s configuration nor critical handling of files. It doesn’t even require access to the server, just a user with the proper permissions to run “show full processlist”. But, and this is a significantly big “but,” you have to take into account that polling the SHOW PROCESSLIST command misses quite a number of queries and gives very poor timing resolution, among other things (like the processlist Mutex).

What’s the alternative? Use the Performance Schema. Specifically: The events_statements_* and threads tables.

First, we have to make sure that we have the correspondent consumers enabled:

Additionally, for statistics to be collected for statements, it is not sufficient to enable only the final statement/sql/* instruments used for individual statement types. The abstract statement/abstract/* instruments must be enabled as well. This should not normally be an issue because all statement instruments are enabled by default.

If you can’t see the event_statements_* consumers on your setup_consumers tables, you’re probably running a MySQL version prior to 5.6.3. Before that version, the events_statements_* tables didn’t exists. MySQL 5.6 might not be more widely used, as was already pointed out in this same blog.

Before continuing, it’s important to note that the most important condition at the moment of capture data is that:

The statements must have ended.

If the statement is still being executed, it can’t be part of the collected traffic. For the ones out there that want to know what’s running inside MySQL, there’s already a detailed non-blocking processlist view to replace [INFORMATION_SCHEMA. | SHOW FULL] PROCESSLIST available with Sys Schema (that will come as default in MySQL 5.7).

Our options to capture data are: get it from one of the three available tables: events_statements_current, events_statements_history or events_statements_history_long.

First option: use the events_statements_current table, which contains current statement events. Since we only want to get statements that have ended, the query will need to add the condition END_EVENT_ID IS NOT NULL to the query. This column is set to NULL when the event starts and updated to the thread current event number when the event ends, but when testing, there were too many missing queries. This is probably because between iterations, the associated threads were removed from the threads table or simply because the time between END_EVENT_ID being updated and the row being removed from the table is too short. This option is discarded.

Second option: The events_statements_history table contains the most recent statement events per thread and since statement events are not added to the events_statements_history table until they have ended, using this table will do the trick without additional conditions in order to know if the event is still running or not. Also, as new events are added, older events are discarded if the table is full.

That means that this table size is fixed. You can change the table size by modifying the variable performance_schema_events_statements_history_size. In the server version I used (5.6.25-73.1-log Percona Server (GPL), Release 73.1, Revision 07b797f) the table size is, by default, defined as autosized (-1) and can have 10 rows per thread. For example: if you are running 5 threads, the table will have 50 rows.

Since it is a fixed size, chances are that some events might be lost between iterations.

Third option: The events_statements_history_long table, which is kind of an extended version of events_statements_history table. Depending on the MySQL version, by default it can hold up to 10000 rows or be autosized (also modifiable with the variable performance_schema_events_statements_history_long_size)

One major -and not cool at all- drawback for this table is that “When a thread ends, its rows are removed from the table”. So it is not history-history data. It will go as far as the oldest thread, with the older event still alive.

The logical option to choose would be the third one: use the events_statements_history_long table. I’ve created a small script (available here) to collect infinite iterations on all the events per thread between a range of event_id’s. The idea of the range is to avoid capturing the same event more than once. Turns out that the execute a query against this table is pretty slow, something between 0.53 seconds and 1.96 seconds. It can behave in a quite invasive way.

Which leave us with the second option: The events_statements_history table.

Since the goal is to capture data in a slow log format manner,  additional information needs to be obtained from the threads table, which has a row for each server thread. The most important thing to remember: access to threads does not require a mutex and has minimal impact on server performance.

Combined, these two tables give us enough information to simulate a very comprehensive slow log format. We just need the proper query:

The idea of this query is to get a Slow Log format as close as possible to the one that can be obtained by using all the options from the log_slow_filter variable.

The other conditions are:

  • t.TYPE = ‘FOREGROUND’: The threads table provides information about background threads, which we don’t intend to analyze. User connection threads are foreground threads.
  • t.PROCESSLIST_INFO IS NOT NULL: This field is NULL if the thread is not executing any statement.
  • t.PROCESSLIST_ID != connection_id(): Ignore me (this query).

The output of the query will look like a proper Slow Log output:

And this file can be used with pt-query-digest to aggregate similar queries, just as it was a regular slow log output. I ran a small test which consists of:

  • Generate traffic using sysbench. This is the sysbench command used:
  • Capture the data using slow log + long_query_time = 0
  • Capture data using pt-query-digest –processlist
  • Capture data from Performance Schema
  • Run pt-query-digest on the 3 files

The results were:

– Slow Log:

– pt-query-digest –processlist

– Performance Schema

The P_S data is closer to the Slow Log one than the captured with regular SHOW FULL PROCESSLIST, but it is still far from being accurate. Remember that this is an alternative for a fast and easy way to capture traffic without too much trouble, so that’s a trade-off that you might have to accept.

Summary: Capture traffic always comes with a tradeoff, but if you’re willing to sacrifice accuracy it can be done with minimal impact on server performance, using the Performance Schema. Because P_S is enabled by default since MySQL 5.6.6 you might already be living with the overhead (if using 5.6). If you are one of the lucky ones that have P_S on production, don’t be afraid to use it. There’s a lot of data already in there.

PREVIOUS POST
NEXT POST
Daniel Guzmán Burgos

Daniel studied Electronic Engineering, but quickly becomes interested in all data things. He has worked as a DBA since 2007 for several companies including one of the major social gaming company at that time in Latin America. Daniel lives in Bogotá, Colombia and in his free time he loves to spend time with his family and go out to eat at any kind of restaurant. Working for Percona since 2014, he is part of the Consulting team.

14 Comments

  • Hi Daniel,

    I wish you provided another example than to retrieve something similar to that of the slow query log. This query obviously will add some overhead and may not run in case the server is on its way to crashing. I think to get slow queries, the slow query log is by far the best option. If you need the slow queries in your application, then you can just read that file.

    Also phpMyAdmin offers excellent performance statistics (along with practical recommendations) – I’m not sure why anyone would need to just do all this work himself.

    Please don’t take me wrong, I think these are excellent tips, but I’m just wondering whether it’s worth it to use them in the practical world!

  • This seems like a really convoluted and more lossy method for:

    SELECT .. FROM performance_schema.events_statements_summary_by_digest ORDER BY sum_timer_wait desc;

    I can’t see why you would want to do the above, if you are not ultimately interested in the raw literal values on a per statement basis and only want aggregate data..

  • Hi Daniel,

    Interesting post, and always informative. I thought I would share this with you, as I got around a lot of the issues you have stated with mysql-proxy as a part of the solution I have detailed it here (and this is all open source… and not personal promotion.. just trying to hekp out) : http://woodygsd.blogspot.co.uk/2014/07/how-do-you-log-problem-like-mysql.html

    I realise proxy is not ‘released’… but it works. Even if you run one proxy per server, just to enable this logging to happen. Essentially I wrote some custom Lua code that attaches to proxy. This gets written out to a file, then logstash pushes that to Elasticsearch, allowng Kibana to graph it. Sounds like a huge stack.. it isn’t.

    Currenty he whole is in production with Avaaz (www.avaaz.org) and tracking all queries/connections through 9 servers, amounting to around 120m data items per day.

    Using this I have been able to save 60m queries per day, moved 40m connections off a master onto the slaves, and found out the reasons why the DB was dragging the site down in high load, and stopped it from happening.
    Essentially this system is logging every single query, and not stalling MySql one bit (ok… maybe a millisecond or two.. but nothing more.. and it is worth the extra just to get te information out). We are also able to get actual slow queries, queries by the hour/day/month… alll beautifully aggregated.
    Just in the process of adding web-server logs.. so when things go awry we can actually see which web calls are causing it…. even though I say so myself.. this is way cool.

    I invite you to take a look.. its using pure OpenSource code, so is free to all… just want to help the ‘struggling’ 🙂

    Peter Colclough

  • Hi Fadi El-Eter,

    Absolutely! the slow log is one of the greatest options to capture traffic, but as described in the blog post, under certain circumstances it can hurt the overall performance. Think about a slave whose buffer pool is keep it warm by reproducing the read traffic from the master, something that you can do with Percona Playback https://www.percona.com/blog/2012/10/25/replaying-database-load-with-percona-playback/, instead of having running the slow_query_log with long_query_time=0 all the time (being a potential bottleneck on high concurrency, with a bunch od transactions in “cleaning up” state), you can use this alternative. It’s a scenario where you don’t need 100% the exactly same traffic.

    Another example, less complicated, is track write traffic to a single table. Instead of using the slow log or the binlog files with mysqlbinlog+some filtering, you can get that data directly from this table.

    Hope it helps!

  • Mark,

    My intention when choosing to use pt-query-digest was to show how close to reality (and by reality i mean: “the traffic captured by the slow log file”) was the traffic collected through performance schema. This post only shows an alternative that could be useful in scenarios where you don’t have access to the server and only a user with grants to read P_S, for say one scenario.

    And indeed! for summary purpose, the events_statements_summary_by_digest is perfect and -as long as there’s enough rows on the events_statements_history_long table- you probably can have more than the digest with the placeholders. But also, you probably won’t, which will make the query analysis harder, as pointed some time ago in https://www.percona.com/blog/2014/02/11/performance_schema-vs-slow-query-log/ However, still very useful!

    P.S: Thanks for Sys Schema!

  • Hi Peter,

    Thanks for sharing! I’m always happy to see different alternatives to solve a common problem. I guess this solution falls in the same category as the tcpdump packet capture, since mysql proxy also hooks from the TCP to forwards packets using MySQL network protocol. Great project and very well documented as i see on the Github repo. Now, i wonder: How does mysql-proxy behave under a high concurrency situation? The latency increase while the threads_running increase in a acceptable ratio? How do you monitor the proxy itself? Can you get the exactly same info from P_S?

  • You can also capture traffic using events_statements_summary_by_digest, but you will need a little help.

    This is where Prometheus[0] mysqld_exporter[1] come in.

    With Prometheus, we can collect metrics over time and store them outside of MySQL for easy analysis. The events statements collector stores separate timeseries for the number of queries, the time used by the queries, the rows examined, sent, etc.

    See this tweet for an example of the detail you get. https://twitter.com/matthiasr/status/647369742714576896

    You can quickly answer “which queries are the slowest”, “which queries examine the most rows”.

    There are a bunch of advantages to using timeseries data for monitoring
    * We can also look back through history.
    * We just collect data, we don’t need to answer all the questions ahead of time
    * Prometheus doesn’t down-sample, so you have full resolution metrics for as long as you keep history.
    * You can write alerting rules against the timeseries.

    Here’s a simple example of an alerting rule we use:
    ALERT ThreadCacheMissRateTooHigh
    IF rate(mysql_global_status_threads_created[5m]) / (rate(mysql_global_status_connections[5m]) > 0) > 0.5 FOR 60m

    We used to try and do this with a nagios plugin, but the plugin only looked at the current value of the counters. By examining the timeseries rate over a few min, we get a much more correct look at the state of the hit rate.

    I am currently monitoring about 150 MySQL servers grouped into about 20 different clusters of masters/slaves/xtradb-clusters. This generates around 700k different metrics timeseries at 15s resolution. This is no problem for a single server running Prometheus. We use this to drive all of our status dashboards.

    [0]: http://prometheus.io/
    [1]: https://github.com/prometheus/mysqld_exporter

  • Hrm, somehow my post got eaten after I submitted it.

    There is a much better way to understand what’s going on inside your server. Collect events_statements_summary_by_digest and store them in a timeseries Database.

    I recently completed adding this functionality to the Prometheus[0] mysqld_exporter[1]

    Using this configuration I am collecting stats for every unique query digest across our fleet of 150+ percona servers, covering 25 different clusters with 50 different application services.

    Here’s a great example of how we were able to upgrade from Ganglia mysql stats to prometheus metrics:

    https://twitter.com/matthiasr/status/647369742714576896

    The first graph comes from basic query counts data from SHOW GLOBAL STATUS, the second one is detailed per query stats. We can generate more details on the number of queries, the query latency, the number of rows examined per query, rows sent per query, etc, etc.

    [0]: http://prometheus.io/
    [1]: https://github.com/prometheus/mysqld_exporter

  • Third time’s a charm, my posts seem to end up in /dev/null

    There is a much better way to see what’s going on inside MySQL with the performance schema.

    Prometheus[0] mysqld_exporter[1] can collect metrics from events_statements_summary_by_digest and allow you to analysis on the timeseries data.

    We are monitoring about 150 percona mysql servers setup into about 25 different service clusters. One Prometheus server is able to monitor over 700k timeseries metrics and allow you to query, graph, and alert on this data in real-time.

    Here’s an example of what we were graphing in Ganglia, and now what we can get from Prometheus and performance schema.

    https://twitter.com/matthiasr/status/647369742714576896

    [0]: http://prometheus.io/
    [1]: https://github.com/prometheus/mysqld_exporter

  • An easier alternative for capturing the queries off the wire traffic (without a man-in-the-middle like a proxy) is VividCortex’s traffic analyzer. https://www.vividcortex.com/resources/network-analyzer-for-mysql/

  • The downside of VividCortex is that it doesn’t know anything about what’s going on inside MySQL.

    Performance Schema tells you a lot more information. For example with performance_schema table_io_waits_summary_by_table and events_statements_summary_by_digest I was easily able to discover a performance problem where a query pattern was causing large amounts of memory temp tables to be used.

    Perf schema is also a lot less system overhead, since you don’t need to attempt to pcap everything the server is doing.

  • Percona benchmarked VividCortex’s overhead versus the Performance Schema a few weeks ago. Using libpcap was not a “lot” more overhead (unless perhaps you do it blindly instead of pushing a packet filter into the kernel to capture only the packets needed, which VividCortex does). Crucially, however, they benchmarked with Performance Schema _idle_ and that is not realistic – in reality you’re going to be querying it frequently as shown in this blog post, and that will add overhead. In general, VividCortex is not higher overhead than P_S, there is just no free lunch.

    Your example of finding queries that use large amounts of memory temp tables is good, but we can do the same thing with VividCortex. In fact we’re not limited to the things that are instrumented in Performance Schema, we can do it for arbitrary metrics, which need not even be from within MySQL! See this blog post that explains the principle at work (although obliquely): https://www.vividcortex.com/blog/2015/10/15/end-to-end-traces/

  • Hello, I am using Workbench 6.3.9 and I can’t see prepared statements using PerformanceReports->HighCostSqlStatements->StatementsAnalysis. Is this feature or on purpose?
    Jakub

Leave a Reply