(The Curious Case of… used to be part of our bi-weekly newsletter but we decided to make it a regular blog post instead so it can sometimes be more frequent. It covers something interesting one of us encountered when working with a client, doing some testing, or were asked in a random question from the community.)
And it’s an auspicious day to be posting, as SQLskills/SYSolutions turns 24 today on 7/24/19!
I had an email from someone last week asking why the transaction log has to be zero-initialized and a I realized I’ve never written about this before, so here goes. (Note: I do explain and demonstrate this in my Pluralsight course on SQL Server: Logging, Recovery, and the Transaction Log.)
It’s all to do with crash recovery. SQL Server knows where crash recovery has to start for a database, but not where it ends – i.e. SQL Server does not persist the ‘most recent LSN’ for a database anywhere. This means it has to work out where the end of the log is (and by end, I mean the most recent log record persisted on disk, not the physical end of the log file).
Some background before the explanation:
- The log is split up internally into chunks called virtual log files, or more commonly just VLFs.
- The first time a VLF is activated and used, all used sections of it are stamped with parity bits 64 (the mechanism for this is not important)
- Eventually the VLF will be marked inactive, and eventually reused. The second time a VLF is activated, all used sections of it are stamped with parity bits 128
- And then 64 again
- And then 128 again
- Ad infinitum…
Why 64 and 128 as the alternating parity bits you may ask? Why not? is my response. I can’t think of a reason to use a different bit pattern pair.
Back to the question at hand…
The most common case when crash recovery happens is that the log has wrapped around a few times and so the various VLFs have been activated and deactivated a few times. Crash recovery goes to where it must start: either the most recent checkpoint, or the start of the oldest active transaction at the time the most recent checkpoint happened. It follows the sequence of active VLFs until it comes to a point where a section of a VLF has the wrong parity bits. This means a VLF is active and says all valid sections should have parity bits X, and crash recovery finds an old section of the VLF from its previous use that has parity bits Y. That’s the end of the log. (Yes, when a VLF is reactivated, it is not zeroed out, because the overwriting of the old sections with new sections with new parity bits works instead.)
The much rarer case is when the log *hasn’t* yet wrapped around and not all the VLFs in the log have been used. In that case, crash recovery proceeds from the start until it find a section of an active VLF that is full of zeroes. And that’s the end of the log in that case.
New physical portions of the log file have to zero-initialized as the previous bits and bytes on disk might just happen to look like a section of a VLF with the ‘correct’ parity bits, causing crash recovery to try to use it and most likely crash SQL Server. It’s highly improbable, but there’s a very small possibility.
Hence the log cannot use instant file initialization.
PS: note that in SQL Server 2016, the ‘zeroing’ pattern changed from 0x00 (hexadecimal zero) to 0xc0, for reasons unrelated to what we’re discussing here.