Reconsidering access paths for index ordering… a dangerous optimization… and a fix!

MySQL has had an interesting optimization for years now1, which has popped up from time to time: in certain circumstances, it may choose to use an index that is index-wise less efficient, but provides the resulting rows in order, to avoid a filesort of the result.

What does this typically look like in production? A query that seems simple and easy takes much longer than it should, sometimes. (Perhaps in production, the query gets killed by pt-kill or exceeds the max_execution_time provided.) The query could be very simple indeed:

SELECT ... WHERE `other_id` = 555 ORDER BY `id` ASC LIMIT 1

There’s an index on other_id, and running the query with an appropriate USE INDEX, the query is fast. Even weirder, changing the query to use LIMIT 10 causes it to run lightning-fast! If anything, the LIMIT 1 should be faster… so, what gives?

Looking at the EXPLAIN, you may notice that the LIMIT 1 version is using access type index on the PRIMARY key (in this case id), whereas the LIMIT 10 version is using ref on a secondary key. Access type index means a full-index scan… on the primary key… which is the clustered key… which is a full table scan.

The optimization is hoping that the LIMIT n with a small enough limit will allow execution to be completed early, without scanning many rows, once the LIMIT is satisfied. This hope is often misplaced: there is no guarantee that there will be any matching rows in the first m rows of the table when ordered by the unwisely-chosen index. Hope is not a strategy.

Although the underlying issue had been reported several times already, under various circumstances, since there were so many individual bugs reported, I filed a new bug with a summary of the situation… and a patch: MySQL Bug 97001.

On MySQL Bug 97001, I proposed a solution (and provided a patch, submitted to Oracle under the NDA) introducing a new optimizer_switch flag named reconsider_index_for_order (defaulting to on to duplicate the current behavior). Although this optimization might benefit some queries, it’s too dependent on the actual data in the involved tables, so it’s not a good candidate for a general optimizer feature in my opinion. Maybe the default could eventually be off allowing users to opt into this optimization when they want it but providing compatible default behavior.

1 The underlying optimization actually appears to be have roots in some very old code, most importantly probably one specific commit in 5.1.

Thanks, Oracle, for fixing the stupid and dangerous SET GLOBAL sql_log_bin!

As of MySQL 5.5.41, released on November 28 (last week), Oracle has fixed MySQL Bug 67433, which I filed on October 31, 2012 and wrote about 4 months ago in Stupid and dangerous: SET GLOBAL sql_log_bin. You can see the fix in Launchpad revision 4718.

The MySQL 5.5.41 release notes mention:

Replication: The global scope for the sql_log_bin system variable has been deprecated, and this variable can now be set with session scope only. The statement SET GLOBAL SQL_LOG_BIN now produces an error. It remains possible for now to read the global value of sql_log_bin, but you should act to remove from your applications any dependencies on reading this value, as the ability to do so will be removed in a future MySQL release. (Bug #67433, Bug #15868071)

Thanks for addressing this and making the world’s data (and DBAs’ jobs) a little safer, Oracle! Special thanks to Morgan Tocker (MySQL Community Manager at Oracle) who cared about it enough to keep pushing it through to a reasonable fix/resolution.

Visualizing the impact of ordered vs. random index insertion in InnoDB

[This post refers to innodb_ruby version 0.9.11 as of October 2, 2014.]

Many DBAs know that building indexes in “random” order (or really any order that greatly differs from ordered by key) can be much less efficient. However, it’s often hard to really understand why that is. With the “-illustrate” visualization modes available in innodb_ruby, it’s possible to quite easily visualize the structure of indexes. The space-lsn-age-illustrate mode to innodb_space allows visualization of all pages in a space file by “LSN age”, generating something like a heatmap of the space file based on how recently each page was modified.

(Note that a small Ruby script generate_data_simple.rb was used to generate the test tables used below. You’ll also want to have a reasonable monospace font supporting Unicode block characters correctly, for which I’d heartily recommend Adobe’s Source Code Pro.)

Building an index by insertion in key order

Inserting rows into an index in key order is much more efficient primarily for two reasons:

  1. Pages can be filled completely, and the database (with sufficient optimizations present) can detect the “bulk loading” behavior and efficiently split pages by creating new, empty pages, rather than splitting pages in half.
  2. Only the “edge” of the index is being written to; once a page has filled it will not be revisited. This can make the buffer pool, which caches database pages, much more effective.

Here’s an example of a table with only a PRIMARY KEY clustered index, built by inserting rows in key order:

$ innodb_space -s ibdata1 -T test/t space-lsn-age-illustrate

As you can see in the above image, the index pages are written to in nearly perfect order as they are allocated from the beginning of the file towards the end.

The first extent (the first line of the image) is allocated as a fragment extent, and contains single pages allocated for different purposes. You can clearly see the tablespace bookkeeping pages, FSP_HDR at page 0 and INODE at page 2, immediately followed by the root index page at page 3, all very recently modified. Following that are 32 individually allocated “fragment” pages which were allocated first in the index before it switched to allocating full extents. Then the index pages progress perfectly until the end of the used space.

Also note the LSN age histogram (printed in white above the colored legend at the bottom) shows all buckets equally filled.

Building an index by insertion in random order

Looking at an index built by insertion in key order was reassuring. What if it is built in random order instead? The result is perhaps a bit surprising:

$ innodb_space -s ibdata1 -T test/t_shuffle space-lsn-age-illustrate

Because the rows were inserted in completely random order, every page has an equal chance for insertion. This means in practice that every page is very recently modified, and this is clearly visible by the entire heatmap being purple. What this also means is that the entire table must be continually present in the buffer pool, and if it doesn’t fit, performance will suffer greatly. This is the main reason that performance of insertion in random order is terrible.

Additionally, you can see a few misbehaviors of InnoDB: Note the extents starting at page 1088, 1152, and 1216. Why do you think they look like that? Let me know your ideas in the comments.

Building a primary and secondary index in primary key order

What if you have multiple indexes? Looking at a very simple and typical case, inserting data in key order by the primary key, while a secondary index exists:

$ innodb_space -s ibdata1 -T test/t_index_before space-lsn-age-illustrate

Notice that this looks like the previous cases perfectly interleaved together, because it is exactly that. Since the primary key and secondary index contain completely different data, insertion is ordered by the primary key, but completely mis-ordered by the secondary index, resulting in the secondary index being built inefficiently.

Adding a secondary index to an existing table

The obvious answer then to the previous problem is to add the index after the data has been loaded, and this does in fact result in the expected outcome:

$ innodb_space -s ibdata1 -T test/t_index_after space-lsn-age-illustrate

When the index is built on the existing table (via ALTER TABLE ... ADD INDEX), it is built by scanning and sorting the data before insertion into the index, resulting in an optimal (and very fast) index build.

Visualizing page fill rate

While the above illustrations show how recently each page was modified during the index build, it’s also possible to use space-extents-illustrate to visualize how full each page is. The key-ordered index looks like this:

$ innodb_space -s ibdata1 -T test/t space-extents-illustrate

Compared to the random-ordered index:

$ innodb_space -s ibdata1 -T test/t_shuffle space-extents-illustrate

The random-ordered insertion caused pages to be split more frequently, and in some cases to be severely under-filled, causing a very poor page fill rate on average.

Those of you who are particularly observant will have noticed that the index illustrations in the first pictures above showed that the ordered-insertion index is significantly smaller than the random one. You can see here that the random-insertion index is 41% larger at 1043 pages compared to just 737 pages for the ordered-insertion index. Additionally, 206 more pages are left unused, making the actual disk space usage 57% larger.

Stupid and dangerous: SET GLOBAL sql_log_bin

It’s been almost 4.5 years since, during some code refactoring, it was decided (or accidentally changed?) that sql_log_bin should become a GLOBAL variable as well as a SESSION one. Almost 2 years ago, during MySQL 5.5 upgrades at Twitter, I filed MySQL Bug 67433 describing in detail how stupid and dangerous this change was, and asking for a reversal.

Nothing has been changed or reversed, so SET GLOBAL sql_log_bin continues to be allowed. I continue to directly see, and hear about damage caused by incorrect use of SET GLOBAL sql_log_bin. So, let me describe just how stupid and dangerous it really is.

What really happens when you run SET GLOBAL sql_log_bin?

If you have a master-slave replication configuration, your master will have binary logging enabled (log_bin) and be recording each transaction to its local binary logs, which the slaves read (more or less in real time) in order to replicate. This log must be complete for replication to be useful and trustworthy. If it is incomplete, serious data loss may occur.

It’s been possible for a long time to use SET [SESSION] sql_log_bin=0 within a connection (with a SUPER user) to temporarily disable binary logging of a session. This can be quite useful to be able to run e.g. ALTER TABLE commands without having them immediately replicated to the slaves.

Most people would guess that running the very similar SET GLOBAL sql_log_bin=0 would disable binary logging for all sessions (you know, GLOBAL?) immediately. However, that’s not quite the case.

As background, there are essentially three different scopes of variables in MySQL server:

  • Global-only: The variable only exists, and only makes sense, in a global scope. An example is innodb_buffer_pool_size — there is only one buffer pool, so it makes no sense in any session context. These variables may also be read-only, so that they can’t be changed at runtime.
  • Global and session: The global variable defines the session default, and is copied to the session variable on connect. A user may change the variable within their session if they want, and there may be restrictions on doing so. If the global variable is changed, it will only affect new sessions; existing sessions will keep the variable they copied at session start. A typical example of this is max_allowed_packet.
  • Session-only: There is no global variable associated with the session variable. A user may change the variable only within their session.

The sql_log_bin variable was previously Session-only and with the commit mentioned above, it was made Global and session.

There are a few implications of this:

  • When a session is initiated, it copies its sql_log_bin value from the global scope to its session scope copy.
  • If the DBA issues SET GLOBAL sql_log_bin=0, it only affects new sessions, any already-connected sessions will continue logging. This makes it pretty useless for managing a failover.
  • Changes to the global variable using SET GLOBAL do not even affect the currently connected session; issuing SET GLOBAL sql_log_bin=0 within a session does not prevent the following commands in that session from being binary logged. This is actually often the first sign that something didn’t work as the DBA expected—they suddenly find all slaves to be executing something they thought they had prevented from being replicated.
  • If a connection is made while sql_log_bin=0 and the DBA then issues SET GLOBAL sql_log_bin=1 the sessions which started while it was 0 will continue to not log anything to the binary log for as long as they are allowed to stay connected. This also makes it pretty useless for managing a failover.
  • If the DBA accidentally issues SET GLOBAL sql_log_bin=0 briefly, even for milliseconds, the entire master-slave replication topology may be irreparably damaged and if this is done without immediately noticing it, permanent and severe data loss may occur.

Is anyone putting it to good use?

The only comment that MySQL Bug 67433 has received from Oracle was from Sveta Smirnova, saying:

Regarding to “revert” I don’t agree: with global access it is possible to make a server master online, without bringing it down.

If you Google search for “set global sql_log_bin” you can find plenty of examples of people talking about it in various languages. I couldn’t find a single example of a script or person using or describing its use correctly, and I looked through pages and pages of examples.

So here’s a Hall of Shame instead:

Also notably, the MySQL manual text for sql_log_bin also still does not adequately describe or warn about its behavior.

Your move, Oracle

Do the right thing, and revert this change. We even provided a patch.

Update: Morgan Tocker, MySQL Community Manager at Oracle is requesting feedback on the future of SET GLOBAL sql_log_bin. Go and make your opinion known!

Thanks, Oracle for fixing MySQL Bug 69477

Just under a year ago, on June 14, 2013, I reported MySQL Bug 69477 (and MariaDB Bug MDEV-4662) titled “InnoDB: Use of large externally-stored fields makes crash recovery lose data”. The gist of this bug is that if you use large BLOB fields (> ~10% of your InnoDB redo logs size), you can compromise crash recovery and suffer data loss after a crash.

The MySQL bug was marked as “Verified” in less than an hour (wow!), but since then there was not a single word about it. Suddenly a couple of days ago, I got an email with “[Com,Ver->Csd]” in the subject. My immediate thought was “Damn it, closed?! How can they close it?! It’s a bug!” — however, upon opening the email I noticed it was not closed as “Won’t fix”, but rather, it was fixed!

Daniel Price at Oracle left the following comment on the bug:

Fixed as of 5.6.20, 5.7.5, and here’s the changelog entry:

Redo log writes for large, externally stored “BLOB” fields could overwrite the most recent checkpoint. The 5.6.20 patch limits the size of redo log “BLOB” writes to 10% of the redo log file size. The 5.7.5 patch addresses the bug without imposing a limitation. For MySQL 5.5, the bug remains a known limitation.

The limitation has also been documented in the reference manual. The changes should appear soon, with the next published documentation build.

Thank you for the bug report.

Thanks, Oracle, for fixing1 this bug! I look forward to seeing the fix in both 5.6 and 5.7. I was just thinking the other day that I might have to fix it myself, because it seemed like nobody was working on it.

1 Maybe next time leave a comment that you’re working on it? Pretty please? :)