column compressionVery often, database performance is affected by the inability to cache all the required data in memory. Disk IO, even when using the fastest devices, takes much more time than a memory access. With MySQL/InnoDB, the main memory cache is the InnoDB buffer pool. There are many strategies we can try to fit as much data as possible in the buffer pool, and one of them is data compression.

With regular MySQL, to compress InnoDB data you can either use “Barracuda page compression” or “transparent page compression with punch holes”. The use of the ZFS filesystem is another possibility, but it is external to MySQL and doesn’t help with caching. All these solutions are transparent, but often they also have performance and management implications. If you are using Percona Server for MySQL, you have yet another option, “column compression with dictionary“. This feature is certainly not receiving the attention it merits. I think it is really cool—let me show you why.

We all know what compression means, who has not zipped a file before attaching it to an email? Compression removes redundancy from a file. What about the dictionary? A compression dictionary is a way to seed the compressor with expected patterns, in order to improve the compression ratio. Because you can specify a dictionary, the scope of usefulness of column compression with the Percona Server for MySQL feature is greatly increased. In the following sections, we’ll review the impacts of a good dictionary, and devise a way to create a good one without any guessing.

A simple use case

A compression algorithm needs a minimal amount of data in order to achieve a reasonable compression ratio. Typically, if the object is below a few hundred bytes, there is rarely enough data to have repetitive patterns and when the compression header is added, the compressed data can end up larger than the original.

Compressing a string of three bytes results in a binary object of 15 bytes. That’s counter productive.

In order to illustrate the potential of the dictionary, I used this dataset:

It is a set of 100k Jeopardy questions written in JSON. To load the data in MySQL, I created the following table:

Then, I did some formatting to create insert statements:

And I executed the inserts. About 20% of the rows had some formatting issues but nevertheless, I ended up with close to 80k rows:

The average row length is 316 bytes for a total data size of 23.55MB. The question JSON objects are large enough to matter, but barely large enough for compression. Here are the first five rows:

Let’s begin by a straight column compression without specifying a dictionary:

As expected the data didn’t compress much. The compression ration is 0.82 or if expressed as a percentage, 18%. Since the JSON headers are always the same, and are present in all questions, we should minimally use them for the dictionary. Trying a minimal dictionary made of the headers gives:

There is a little progress, we now have a compression ratio of 0.79. Obviously, we could do more but without a tool, we’ll have to guess. A compressor like zlib builds a dictionary as part of its compression effort, could we use that? Yes, but only if we can generate it correctly and access the result. That’s not readily available with the common compressors I know. Fortunately, someone else had the same issue and wrote a compressor able to save its dictionary. Please let me introduce femtozip.

Femtozip to the rescue

The tool, by itself, has no magic algorithm. It is based on zlib, from what I can understand from the code. Anyway, we won’t compress anything with it, we’ll use it to generate a good dictionary. In order to create a dictionary, the tool looks at a set of files and try to see patterns between them. The use of a single big file defeat the purpose. So, I generated one file per question with:

Then, I used the following command to generate a 1024 bytes dictionary using all the files starting by “1”:

In about 10s the job was done. I tried with all the 80k files and… I had to kill the process after thirty minutes. Anyway, there are 11111 files starting with “1”, a very decent sample. Our generated dictionary looks like:

With some formatting, I was able to create a dictionary with the above data:

That’s interesting, we are now achieving a ratio of 0.61, a significant improvement. I pushed my luck and tried with a 2048 bytes dictionary. It further reduced the ratio to 0.57 but that was about the best I got. Larger dictionaries didn’t lower the ratio below 0.57. Zlib supports up to 32KB for the dictionary.

So, to recap:

  • column compression without dictionary, ratio of 0.82
  • column compression with simple dictionary, ratio of 0.79
  • column compression with a 1k dictionary from femtozip, ratio of 0.61
  • column compression with a 2k dictionary from femtozip, ratio of 0.57

The above example stores a JSON document in a text column. MySQL 5.7 includes a JSON datatype which behaves a bit differently regarding the dictionary. Delimiting characters like ‘{}’ are removed in the on disk representation of a JSON column. If you have TBs of data in similar tables, you should really consider column compression and a systematic way of determining the dictionary with femtozip. In addition to improve the compression, it is likely to be the less performance impacting solution. Would it be interesting to generate a dictionary from existing data with a command like this one?

where the dictionary creation process would implicitly includes steps similar to the ones I did with femtozip.

2 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
Jouni Järvinen

You sure you need to create a separate process for running the §while§ loops ?

0
Yves Trudeau

@Jouni It does work… Did I write it has to be like that? Just my Bash coding habits.

0