Tag Archives: InnoDB

Improving InnoDB index statistics

The MySQL/MariaDB optimiser likes to know things like the cardinality of an index – that is, the number of distinct values the index holds. For a PRIMARY KEY, which only has unique values, the number is the same as the number of rows.  For an indexed column that is boolean (such as yes/no) the cardinality would be 2.

There’s more to it than that, but the point is that the optimiser needs some statistics from indexes in order to try and make somewhat sane decisions about which index to use for a particular query. The statistics also need to be updated when a significant number of rows have been added, deleted, or modified.

In MyISAM, ANALYZE TABLE does a tablescan where everything is tallied, and the index stats are updated. InnoDB, on the other hand, has always done “index dives”, looking at a small sample and deriving from that. That can be ok as a methodology, but unfortunately the history is awkward. The number used to be a constant in the code (4), and that is inadequate for larger tables. Later the number was made a server variable innodb_stats_sample_pages and its default is now 8 – but that’s still really not enough for big(ger) tables.

We recently encountered this issue again with a client, and this time it really needed addressing as no workarounds were effective across the number of servers and of course over time. Open Query engineer Daniel filed https://mariadb.atlassian.net/browse/MDEV-7084 which was picked up by MariaDB developed Jan Lindström.

Why not just set the innodb_stats_sample_pages much higher? Well, every operation takes time, so setting the number appropriate for your biggest table means that the sampling would take unnecessarily long for all the other (smaller, or even much smaller) tables. And that’s just unfortunate.

So why doesn’t InnoDB just scale the sample size along with the table size? Because, historically, it didn’t know the table size: InnoDB does not maintain a row count (this has to do with its multi-versioned architecture and other practicalities – as with everything, it’s a trade-off). However, these days we have persistent stats tables – rather than redoing the stats the first time a table is opened after server restart, they’re stored in a table. Good improvement. As part of that information, InnoDB now also knows how many index pages (and leaf nodes in its B+Tree) it has for each table. And while that’s not the same as a row count (rows have a variable length so there’s no fixed number of rows per index page), at least it grows along with the table. So now we have something to work with! The historical situation is no longer a hindrance.

In order to scale the sample size sanely, that is not have either too large a number for small tables, or a number for big tables that’s over the top, we’ll want some kind of logarithmic scale. For instance, log2(16 thousand) = 14, and log2(1 billion) = 30. That’s small enough to be workable. The new code as I suggested:

n_sample_pages = max(min(srv_stats_sample_pages, index->stat_index_size), log2(index->stat_index_size) * srv_stats_sample_pages);

This is a shorter construct (using min/max instead of ifs) of what was already there, combined with the logarithmic sample basis. For very small tables, either the innodb_stats_sample_pages number if used or the actual number of pages, whichever is smaller – for bigger tables, the log2 of the #indexpages is used, multiplied by the dynamic system variable innodb_stats_sample_pages. So we can still scale and thus influence the system in case we want more samples. Simple, but it seems effective – and it any case we get decent results in many more cases than before, so it’s a worthwhile improvement. Obviously, since it’s a statistical sample, it could still be wrong for an individual case.

Jan reckons that just like MyISAM, InnoDB should do a table scan and work things out properly – I agree, this makes sense now that we have persistent stats. So the above is a good fix for 5.5 and 10.0, and the more significant change to comprehensive stats can be in a near future major release. So then we have done away with the sampling altogether, instead basing the info on the full dataset. Excellent.

Another issue that needed to be dealt with is when InnoDB recalculates the statistics. You don’t want to do it on every change, but regularly if there has been some change is good as it might affect which indexes should be chosen for optimal query execution. The hardcoded rule was 1/16th of the table or 2 billion rows, whichever comes first. Again that’s unfortunate, because for a bigger table 1/16th still amounts to a very significant number. I’d tend towards setting the upper bound to say 100,000. Jan put in a new dynamic system variable for this, stat_modified_counter. It essentially replaces the old static value of 2 billion, providing us with a configurable upper bound. That should do nicely!

Once again horay for open source, but in particular responsive and open development. If someone reports a bug and there is no interaction between developer and the outside world until something is released, the discussion and iterative improvement process to come to a good solution cannot occur. The original code definitely worked as designed, but it was no longer suitable in the context of today’s usage/needs. The user-dev interaction allowed for a much better conclusion.

innodb_flush_logs_on_trx_commit and Galera Cluster

We deploy Galera Cluster (in MariaDB) for some clients, and innodb_flush_logs_on_trx_commit is one of the settings we’ve been playing with. The options according to the manual:

  • =0 don’t write or flush at commit, write and flush once per second
  • =1 write and flush at trx commit
  • =2 write log, but only flush once per second

The flush (fsync) refers to the mechanism the filesystem uses to try and guarantee that written data is actually on the physical medium/device and not just in a buffer (of course cached RAID controllers, SANs and other devices use some different logic there, but it’s definitely written beyond the OS space).

In a non-cluster setup, you’d always want it to be =1 in order to be ACID compliant and that’s also InnoDB’s default. So far so good. For cluster setups, you could be more lenient with this as you require ACID on the cluster as a whole, not each individual machine – after all, if one machine drops out at any point, you don’t lose any data.

Codership docu recommended =2, so that’s what Open Query engineer Peter Lock initially used for some tests that he was conducting. However, performance wasn’t particularly shiny – actually not much higher than =1. That in itself is interesting, because typically we regard the # of fsyncs/second a storage system can deal with as a key indicator of performance capacity. That is, as our HD Latency tool shows when you run it on a storage device (even your local laptop harddisk), the most prominent aspect of what limits the # of writes you can do per second appears to be the fsyncs.

I then happened to chat with Oli Sennhauser (former colleague from MySQL AB) who now runs the FromDual MySQL/MariaDB consulting firm in Switzerland, and he’s been working with Galera for quite a long time. He recognised the pattern and said that he too had that experience, and he thought =0 might be the better option.

I delved into the InnoDB source code to see what was actually happening, and the code indeed concurs with what’s described in the manual (that hasn’t always been the case ;-). I also verified this with Jeremy Cole whom we may happily regard as guru on “how InnoDB actually works”. The once-per-second flush (and optional preceding write) is performed by the InnoDB master thread. Take a peek in log/log0log.c and trx/trx0trx.c, specifically trx_commit_off_kernel() and srv_sync_log_buffer_in_background().

In conclusion:

  1. Even with =0, the log does get written and flushed once per second. This is done in the background so connection threads don’t have to wait for it.
  2. There is no setting where there is never a flush/fsync.
  3. With =2, the writing of the log takes place in the connection thread and this appears to incur a significant overhead, at least relative to =0. Aside from the writing of the log at transaction commit, there doesn’t appear to be a difference.
  4. Based on the preceding points, I would say that if you don’t want =1, you might as well set =0 in order to get the performance you’re after. There is of course a slight practical difference between =0 and =2. With =2 the log is immediately written. If the mysqld process were to crash within a second after that, the OS would close the file and have that log write stored. With =0 that log data wouldn’t have been written. If the OS or machine fails, that log write is lost either way.

In production environments, we tend to mainly want to mitigate trouble from system failures, so =0 appears to be a suitable/appropriate option – for a Galera cluster environment.

What remains is the question of why the log write operation appears to reduce transaction commit performance so much, in a way more so than the flush/fsync. Something to investigate further!
Your thoughts?

Galera pre-deployment check

One of the first things we do when preparing a client’s infrastructure for Galera deployment is see whether their schema is suitable.

  • Avoiding quirks and edge cases, we can say that Galera simply requires all tables to be InnoDB and also have a PRIMARY KEY (obviously having a PK in InnoDB is important anyway, for InnoDB-internal reasons).
  • We want to know about FULLTEXT indexes. With recent InnoDB versions also supporting FULLTEXT we need to check not just whether a table has such an index, but actually which engine it is.
  • Spatial indexes. While both InnoDB and MyISAM can deal with spatial datatypes (POINT, GEOMETRY, etc), only MyISAM has the spatial indexes.

Naturally, checking a schema in the server is more effective than going through other sources and possibly missing bits. On the downside, the only viable way to get this info out of MariaDB is INFORMATION_SCHEMA, but because of the way it’s implemented queries tend to be slow and resource intensive. So essentially we do need to ask I_S, but do it as efficiently as possible (we’re dealing with production systems). We have multiple separate questions to ask, which normally we’d ask in separate queries, but in case of I_S that’s really something to avoid. So that’s why it’s all integrated into the single query below, catching every permutation of “not InnoDB”, “lacks primary key”, “has fulltext or spatial index”. We skip the system databases and any VIEWs.

We use the lesser known mysql client command ‘tee’ to output the data into a file, and close it after the query.

We publish the query not as a work of art – I don’t think it’s that pretty! We’d like you to see because we don’t care for secrets and also because if there is any way you can reach the same objective using a less resource intensive approach, we’d love to hear about it! This is one of the very few cases where we care only about efficiency, not how pretty the query looks. That said, of course I’d prefer it to be easily readable.

If you regard it purely as a query to be used for Galera, then you can presume it’ll be run on MariaDB 5.5 or later – since 5.3 and above has optimised subqueries, perhaps you can do something with that.

If you spot any other flaw or missing bit, please comment on that too. thanks!

-- snip
tee galeracheck.txt

SELECT DISTINCT
       CONCAT(t.table_schema,'.',t.table_name) as tbl,
       t.engine,
       IF(ISNULL(c.constraint_name),'NOPK','') AS nopk,
       IF(s.index_type = 'FULLTEXT','FULLTEXT','') as ftidx,
       IF(s.index_type = 'SPATIAL','SPATIAL','') as gisidx
  FROM information_schema.tables AS t
  LEFT JOIN information_schema.key_column_usage AS c
    ON (t.table_schema = c.constraint_schema AND t.table_name = c.table_name
        AND c.constraint_name = 'PRIMARY')
  LEFT JOIN information_schema.statistics AS s
    ON (t.table_schema = s.table_schema AND t.table_name = s.table_name
        AND s.index_type IN ('FULLTEXT','SPATIAL'))
  WHERE t.table_schema NOT IN ('information_schema','performance_schema','mysql')
    AND t.table_type = 'BASE TABLE'
    AND (t.engine <> 'InnoDB' OR c.constraint_name IS NULL OR s.index_type IN ('FULLTEXT','SPATIAL'))
  ORDER BY t.table_schema,t.table_name;

notee
-- snap

Credit: the basis of the “find tables without a PK” is based on SQL by Sheeri Cabral and Giuseppe Maxia.

Understanding SHOW VARIABLES: DISABLED and NO values

When you use SHOW VARIABLES LIKE “have_%” to see whether a particular feature is enabled, you will note the value of NO for some, and DISABLED for others. These values are not intrinsically clear for the casual onlooker, and often cause confusion. Typically, this happens with SSL and InnoDB. So, here is a quick clarification!

  • NO means that the feature was not enabled (or was actively disabled) in the build. This means the code and any required libraries are not present in the binary.
  • DISABLED means that the feature is built in and capable of working in the binary, but is disabled due to relevant my.cnf settings.
  • YES means the feature is available, and configured in my.cnf.

SSL tends to show up as DISABLED, until you configure the appropriate settings to use it in my.cnf (SHOW VARIABLES LIKE “ssl_%”). From then on it will show up as YES.

Depending on your MySQL version and distro build, InnoDB can be disabled via the “skip-innodb” option. Obviously that’s not recommended as InnoDB should generally be your primary engine of choice!

However, InnoDB can also show up as DISABLED if the plugin fails to load due to configuration or other errors on startup. When this happens, review the error log (often redirected to syslog/messages) to identify the problem.

If InnoDB is configured as the default storage engine, failed initialisation of the plugin should now result in mysqld not starting, rather than starting with InnoDB disabled, as obviously InnoDB is required in that case.

What a Hosting Provider did Today

I found Dennis the Menace, he now has a job as system administrator for a hosting company. Scenario: client has a problem with a server becoming unavailable (cause unknown) and has it restarted. MySQL had some page corruption in the InnoDB tablespace.

The hosting provider, being really helpful, goes in as root and first deletes ib_logfile* then ib* in /var/lib/mysql. He later says “I am sorry if I deleted it. I thought I deleted the log only. Sorry again.”  Now this may appear nice, but people who know what they’re doing with MySQL will realise that deleting the iblogfiles actually destroys data also. MySQL of course screams loudly that while it has FRM files it can’t find the tables. No kidding!

Then, while he’s been told to not touch anything any more, and I’m trying to see if I can recover the deleted files on ext3 filesystem (yes there are tools for that), he goes in again and puts an ibdata1 file back. No, not the logfiles – but he had those somewhere else too. The files get restored and turn out to be two months old (no info on how they were made in the first place but that’s minor detail in this grand scheme). All the extra write activity on the partition would’ve also made potential deleted file recovery more difficult or impossible.

This story will still get a “happy” ending, using a recent mysqldump to load a new server at a different hosting provider. Really – some helpfulness is not what you want. Secondary lesson: pick your hosting provider with care. Feel free to ask us for recommendations as we know some excellent providers and have encountered plenty of poor ones.

HDlatency – now with quick option

I’ve done a minor update to the hdlatency tool (get it from Launchpad), it now has a –quick option to have it only do its tests with 16KB blocks rather than a whole range of sizes. This is much quicker, and 16KB is the InnoDB page size so it’s the most relevant for MySQL/MariaDB deployments.

However, I didn’t just remove the other stuff, because it can be very helpful in tracking down problems and putting misconceptions to rest. On SANs (and local RAID of course) you have things like block sizes and stripe sizes, and opinions on what might be faster. Interestingly, the real world doesn’t always agree with the opinions.

We Mark Callaghan correctly pointed out when I first published it, hdlatency does not provide anything new in terms of functionality, the db IO tests of sysbench cover it all. A key advantage of hdlatency is that it doesn’t have any dependencies, it’s a small single piece of C code that’ll compile on or can run on very minimalistic environments. We often don’t control what the base environment we have to work on is, so that’s why hdlatency was initially written. It’s just a quick little tool that does the job.

We find hdlatency particularly useful for comparing environments, primarily at the same client. For instance, the client might consider moving from one storage solution to another – well, in that case it’s useful to know whether we can expect an actual performance benefit.

The burst data rate (big sequential read or write) which often gets quoted for a SAN or even an individual disk is of little interest to database use, since its key performance bottleneck lies in random access I/O. The disk head(s) will need to move. So it’s important to get some real relevant numbers, rather than just go with magic vendor numbers that are not really relevant to you. Also, you can have a fast storage system attached via a slow interface, and consequentially the performance then will not be at all what you’d want to see. It can be quite bad.

To get an absolute baseline on what are sane numbers, run hdlatency also on a local desktop HD. This may seem odd, but you might well encounter storage systems that show a lower performance than that. ‘nuf said.

If you’re willing to share, I’d be quite interested in seeing some (–quick) output data from you – just make sure you tell what storage it is: type of interface, etc. Simply drop it in a comment to this post, so it can benefit more people. thanks

Fast paging in the real world

This blag was originally posted at http://cafuego.net/2010/05/26/fast-paging-real-world

Some time ago I attended the “Optimisation by Design” course from Open Query¹. In it, Arjen teaches how writing better queries and schemas can make your database access much faster (and more reliable). One such way of optimising things is by adding appropriate query hints or flags. These hints are magic strings that control how a server executes a query or how it returns results.

An example of such a hint is SQL_CALC_FOUND_ROWS. You use it in a select query with a LIMIT clause. It instructs the server to select a limited numbers of rows, but also to calculate the total number of rows that would have been returned without the limit clause in place. That total number of rows is stored in a session variable, which can be retrieved via SELECT FOUND_ROWS(); That simply reads the variable and clears it on the server, it doesn’t actually have to look at any table or index data, so it’s very fast.

This is useful when queries are used to generate pages of data where a user can click a specific page number or click previous/next page. In this case you need the total number of rows to determine how many pages you need to generate links for.

The traditional way is to first run a SELECT COUNT(*) query and then select the rows you want, with LIMIT. If you don’t use a WHERE clause in your query, this can be pretty fast on MyISAM, as it has a magic variable that contains the number of rows in a table. On InnoDB however, which is my storage engine of choice, there is no such variable and consequently it’s not pretty fast.

Paging Drupal

At DrupalConSF earlier this year I’d floated the idea of making Drupal 7 use SQL_CALC_FOUND_ROWS in its pager queries. These are queries generated specifically to display paginated lists of content and the API to do this is pretty straightforward. To do it I needed to add query hint support to the MySQL driver. When it turned out that PostgreSQL and Oracle also support query hints though, the aim became adding hint support for all database drivers.

That’s now done, though only the patch only implements hints on the pager under MySQL at the moment.

One issue keeps cropping up though, a blog by Alexey Kovyrin in 2007 that states SELECT COUNT(*) is faster than using SQL_CALC_FOUND_ROWS. It’s all very well to not have a patch accepted if that statement is correct, but in my experience that is in fact not the case. In my experience the stats are in fact the other way around, SQL_CALC_FOUND_ROWS is nearly always faster than SELECT COUNT(*).

To back up my claims I thought I should run some benchmarks.

I picked the Drupal pager query that lists content (nodes) on the content administration page. It selects node IDs from the node table with a WHERE clause which filters by the content language. Or, in plain SQL, what currently happens is:

SELECT COUNT(*) FROM node WHERE language = 'und';
SELECT nid FROM node WHERE language = 'und' LIMIT 0,50;

and what I’d like to happen is:

SELECT SQL_CALC_FOUND_ROWS nid FROM node WHERE language = 'und' LIMIT 0,50;
SELECT FOUND_ROWS();

Methodology

I ran two sets of tests. One on a node table with 5,000 rows and one with 200,000 rows. For each of these table sizes I ran a pager with 10, 20, 50, 100 and 200 loops, each time increasing the offset by 50; effectively paging through the table. I ran all these using both MyISAM and InnoDB as the storage engine for the node table and I ran them on two machines. One was my desktop, a dual core Athlon X2 5600 with 4Gb of RAM and the other is a single core Xen virtual machine with 512Mb of RAM.

I was hoping to also run tests with 10,000,000 rows, but the virtual machine did not complete any of the queries. So I ran these on my desktop machine only. Again for 10, 20, 50, 100 and 200 queries per run. First with an offset of 50, then with an offset of 10,000. I restarted the MySQL server between each run. To discount query cache advantages, I ran all tests with the query cache disabled. The script I used is attached at the bottom of this post. The calculated times do include the latency of client/server communication, though all tests ran via the local socket connection.

My desktop runs an OurDelta mysql .5.0.87 (the -d10-ourdelta-sail66) to be exact. The virtual machine runs 5.0.87 (-d10-ourdelta65).  Before you complain that not running a vanilla MySQL invalidates the results, I run these because I am able to tweak InnoDB a bit more, so the I/O write load on the virtual machine is somewhat reduced compared to the vanilla MySQL.

Results

Query time graphs - NEW is faster than OLD and InnoDB is not slower than MyISAM

The graphs show that using SQL_CALC_FOUND_ROWS is virtually always faster than running two queries that each need to look at actual data. Even when using MyISAM. As the database gets bigger, the speed advantage of SQL_CALC_FOUND_ROWS increases. At the 10,000,000 row mark, it’s consistently about twice as fast.

Also interesting is that InnoDB seems significantly slower than MyISAM on the shorter runs. I say seems, because (especially with the 10,000,000 row table) the delay is caused by InnoDB first loading the table from disk into its buffer pool. In the spreadsheet you can see the first query takes up to 40 seconds, whilst subsequent ones are much faster. The MyISAM data is still in the OS file cache, so it doesn’t have that delay on the first query. Because I use innodb_flush_method=O_DIRECT, the InnoDB data is not kept in the OS file cache.

Conclusion

So, it’s official. COUNT(*) is dead, long live SQL_CALC_FOUND_ROWS!  :-)

I’ve attached my raw results as a Gnumeric document, so feel free to peruse them. The test script I’ve used is also attached, so you can re-run the benchmark on your own systems if you wish.

Conclusion Addendum

As pointed out in the Drupal pager issue that caused me to run these tests, the query I’m benchmarking uses the language column, which is not indexed and the test also doesn’t allow the server to cache the COUNT(*) query. I’ve rerun the tests with 10 million rows after adding an index and I no longer get a signification speed difference between the two ways of getting the total number of rows.

So I suppose that at least SQL_CALC_FOUND_ROWS will cause your non-indexed pager queries to suck a lot less than they might otherwise and it won’t hurt if they are properly indexed :-)

¹ I now work for Open Query as a consultant.

Unqualified COUNT(*) speed PBXT vs InnoDB

So this is about a SELECT COUNT(*) FROM tblname without a WHERE clause. MyISAM has an optimisation for that since it maintains a rowcount for each table. InnoDB and PBXT can’t do that (at least not easily) because of their multi-versioned nature… different transactions may see a different number of rows for the table table!

So, it’s kinda known but nevertheless often ignored that this operation on InnoDB is costly in terms of time; what InnoDB has to do to figure out the exact number of rows is scan the primary key and just tally. Of course it’s faster if it doesn’t have to read a lot of the blocks from disk (i.e. smaller dataset or a large enough buffer pool).

I was curious about PBXT’s performance on this, and behold it appears to be quite a bit faster! For a table with 50 million rows, PBXT took about 20 minutes whereas the same table in InnoDB took 30 minutes. Interesting!

From those numbers [addendum: yes I do realise there’s something else wrong on that server to take that long, but it’d be slow regardless] you can tell that doing the query at all is not an efficient thing to do, and definitely not something a frontend web page should be doing. Usually you just need a ballpark figure so running the query in a cron job and putting the value into memcached (or just an include file) will work well in such cases.

If you do use a WHERE clause, all engines (including MyISAM) are in the same boat… they might be able to use an index to filter on the conditions – but the bigger the table, the more work it is for the engine. PBXT being faster than InnoDB for this task makes it potentially interesting for reporting purposes as well, where otherwise you might consider using MyISAM – we generally recommend using a separate reporting slave with particular settings anyway (fewer connections but larger session-specific buffers), but it’s good to have extra choices for the task.

(In case you didn’t know, it’s ok for a slave to use a different engine from a master – so you can really make use of that ability for specialised tasks such as reporting.)

PBXT early impressions in production use

With Paul McCullagh’s PBXT storage engine getting integrated into MariaDB 5.1, it’s never been easier to it out. So we have, on a slave off one of our own production systems which gets lots of inserts from our Zabbix monitoring system.

That’s possibly an ideal usage profile, since PBXT is a log based engine (simplistically stated, it indexes its transaction logs, rather than rewriting data from log into index and indexing that) so it should require less disk I/O than say InnoDB. And that means it should be particularly suited to for instance logging, which have lots of inserts on a sustained basis. Note that for short insert burst you may not see a difference with InnoDB because of caching, but sustain it and then you can notice.

Because PBXT has such different/distinct architecture there’s a lot of learning involved. Together with Paul and help from Roland Bouman we also created a stored procedure that can calculate the optimal average row size for PBXT, and even ALTER TABLE statements you can paste to convert tables. The AVG_ROW_LENGTH option is quite critical with PBXT, if set too big (or if you let PBXT guess and it gets it wrong) it’ll eat heaps more diskspace as well as being much slower, and if too small it’ll be slower also; this, it needs to be in the right ballpark. For existing datasets it can be calculated, so that’s what we’ve worked on. The procs will be published shortly, and Paul will also put them in with the rest of the PBXT files.

Another important aspect for PBXT is having sufficient cache memory allocated, otherwise operations can take much much longer. While the exact “cause” is different, one would notice similar performance aspects when using InnoDB on larger datasets and buffers that are too small for the purpose.

So, while using or converting some tables to PBXT takes a bit of consideration, effort and learning, it appears to be dealing with the real world very well so far – and that’s a testament to Paul’s experience. Paul is also very responsive to questions. As we gain more experience, it is our intent to try PBXT for some of our clients that have operational needs that might be a particularly good fit for PBXT.

I should also mention that it is possible to have a consistent transaction between PBXT, InnoDB and the binary log, because of the 2-phase commit (XA) infrastructure. This means that you should even be able to do a mysqldump with –single-transaction if you have both PBXT and InnoDB tables, and acquire a consistent snapshot!

More experiences and details to come.

RAM flakier than expected

Ref: Google: Computer memory flakier than expected (CNET DeepTech, Stephen Shankland)

Summary: According to tests at Google, it appears that today’s RAM modules have several thousand errors a year, which would be correctable if it weren’t for the fact that most of us aren’t using ECC RAM.

Previous research, such as some data from a 300-computer cluster, showed that memory modules had correctable error rates of 200 to 5,000 failures per billion hours of operation. Google, though, found the rate much higher: 25,000 to 75,000 failures per billion hours.

This is quite relevant for database servers because they write a lot rather than mainly read (desktop use). In the MySQL context, if a bit gets flipped in RAM, your data could get corrupted, or it’s ok on disk and you’re just reading corrupted data somehow. While using more RAM is good for performance, it also means a bigger RAM footprint for your data and thus more exposure to the issue.

In MySQL 5.0 and the general 5.1, the binary and relay logs do not have checksums on log events. If something gets corrupted anywhere on disk or on its way to disk, garbage will come out and we have seen instances where this happens. There are patches to add a checksum to the binlog structure (Google worked on this) and we’ll be pushing for this to be ported into MariaDB 5.1 urgently. It’s no use having it just in later versions. It does change the on-disk format, but so be it. This is very very important stuff.

FYI, InnoDB does use page checksums which are also stored on disk. There is an option to turn them off, but our general recommendation would be to not do that ;-) What about the iblog files though? Normally they just refer to pages which at some stage get flushed, but a) if through a glitch they refer to a different page that could lose some committed data and b) on recovery, it could directly affect data. Mind you I’m conjecturing here, more research necessary!

Naturally this does not just affect database systems, file systems too can easily suffer from RAM glitches – probably with the exception of ZFS, since it has checksums everywhere and keeps them separate from the data.

Anything that keeps data around in RAM, and/or is write intensive. Memcached! How do other database systems work in this respect?

Note: this post is not intended to be alarmist; I just think it’s good to be aware of things so they can be taken into account when designing systems. If you look closely at any system, there are things that can potentially be cause for concern. That doesn’t mean we shouldn’t use them, per-say.