I have been working on another PG extension for timeseries (https://github.com/xataio/deltax) for a few months, and trying to score as good as possible on ClickBench.
This is a project that is simply lot of fun to work on. There are many tricks that can be used to speed-up analytics, besides just type-aware compression:
* for each segment you will keep things like max/min/sum, number of distinct values, bloom filters, etc. For a good amount of common queries, you can answer them just based on this metadata, so you don't need to decompress the columns at all.
* for text column, you compress them differently based on cardinality. Low cardinality (think labels or similar) is dictionary based compression. High cardinality is LZ4.
* Generally the smaller the data on disk, the higher the cold runs performance. This is because you need less IO to load it in memory. I have discovered that on top of the type-aware compression, it's worth doing another round of LZ4. There's also some research that it's sometimes worth doing multiple passes of LZ4.
* Partition and segment pruning. If you can tell from the metadata or bloom filters that the filter doesn't match a partition or segment, you skip the whole thing.
* Push down of filters in the decompression layer. Depending on the compression algorithm, while you decompress you can also filter out the values that you don't need. This avoids passing data and allocating memory for elements that will be later discarded anyway.
* Organization of data on disk is more important than almost anything else. Of course, that's the main point of columnar storage, but there are level of details on how to organize the data so that IO is minimized during queries. I have tried 3-4 different layouts before settling on one.
* For top N type of queries, which are really common in analytics, you want to stop the reading from disk / decompressed as soon as you have enough data to guarantee that you have a correct top N to satisfy the query.
* Parallelize everything: at least ClickBench runs on instances with a lot of CPU cores, so you need to parallelize every step of the way. This is done differently depending on the query type. For example for top N, each worker can take a subset of the segments and get the top N from each of them. Then you combine those in a single result.
That section is the most relevant whenever compression in a DB is discussed.
The purpose of a database is to find, aggregate or update data - storage is where the trade-off gets expressed. There are no silver bullets here.
Any method of compression which speeds up either filter rejection or scan rate is better than something that only trades off IO for CPU usage.
For example, dictionary encoding can be slower to read (because you decompress the whole dictionary and not just the skip read after filter), but not if you can squeeze out an IN clause by turning string comparisons into O(1) dictionary followed by a simple integer filter. Remember, this can be arbitrarily complex (Druid is a great example of this) and then the bitmaps can be used because the dictionary index will be a dense 0-N.
Even better if that can feed a deterministic operation like UPPER() so that you do it over the dictionary hits once, instead of each row. You can even use it over the same hash slot, instead of another dictionary collision check or hash computation.
If anyone is looking at JSONB compression, go take a long look at the Variant encoding proposals from Databricks/Snowflake for Iceberg[1].
Turning a single column "payload" JSONB field into chunks which are columnarized and strictly typed allows you to do all the tricks mentioned here, but on loosely typed data but chunk by chunk.
I have been working on another PG extension for timeseries (https://github.com/xataio/deltax) for a few months, and trying to score as good as possible on ClickBench.
This is a project that is simply lot of fun to work on. There are many tricks that can be used to speed-up analytics, besides just type-aware compression:
* for each segment you will keep things like max/min/sum, number of distinct values, bloom filters, etc. For a good amount of common queries, you can answer them just based on this metadata, so you don't need to decompress the columns at all.
* for text column, you compress them differently based on cardinality. Low cardinality (think labels or similar) is dictionary based compression. High cardinality is LZ4.
* Generally the smaller the data on disk, the higher the cold runs performance. This is because you need less IO to load it in memory. I have discovered that on top of the type-aware compression, it's worth doing another round of LZ4. There's also some research that it's sometimes worth doing multiple passes of LZ4.
* Partition and segment pruning. If you can tell from the metadata or bloom filters that the filter doesn't match a partition or segment, you skip the whole thing.
* Push down of filters in the decompression layer. Depending on the compression algorithm, while you decompress you can also filter out the values that you don't need. This avoids passing data and allocating memory for elements that will be later discarded anyway.
* Organization of data on disk is more important than almost anything else. Of course, that's the main point of columnar storage, but there are level of details on how to organize the data so that IO is minimized during queries. I have tried 3-4 different layouts before settling on one.
* For top N type of queries, which are really common in analytics, you want to stop the reading from disk / decompressed as soon as you have enough data to guarantee that you have a correct top N to satisfy the query.
* Parallelize everything: at least ClickBench runs on instances with a lot of CPU cores, so you need to parallelize every step of the way. This is done differently depending on the query type. For example for top N, each worker can take a subset of the segments and get the top N from each of them. Then you combine those in a single result.
> What does compression do to query performance?
That section is the most relevant whenever compression in a DB is discussed.
The purpose of a database is to find, aggregate or update data - storage is where the trade-off gets expressed. There are no silver bullets here.
Any method of compression which speeds up either filter rejection or scan rate is better than something that only trades off IO for CPU usage.
For example, dictionary encoding can be slower to read (because you decompress the whole dictionary and not just the skip read after filter), but not if you can squeeze out an IN clause by turning string comparisons into O(1) dictionary followed by a simple integer filter. Remember, this can be arbitrarily complex (Druid is a great example of this) and then the bitmaps can be used because the dictionary index will be a dense 0-N.
Even better if that can feed a deterministic operation like UPPER() so that you do it over the dictionary hits once, instead of each row. You can even use it over the same hash slot, instead of another dictionary collision check or hash computation.
If anyone is looking at JSONB compression, go take a long look at the Variant encoding proposals from Databricks/Snowflake for Iceberg[1].
Turning a single column "payload" JSONB field into chunks which are columnarized and strictly typed allows you to do all the tricks mentioned here, but on loosely typed data but chunk by chunk.
[1] - https://github.com/apache/parquet-format/blob/master/Variant...
Gorilla by Facebook had this. Value is stored as delta and time as delta of delta.
They say they are using “gorilla compression “
I’m still amazed every time I go back and read how the compression for floating point values works.
It's used in ClickHouse as well. CH supports all known compression algos and they are documented pretty well.