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.
Ha, we (my partner at our company) filed this issue 5 years ago. We have a large-ish (but not giant) json blob in one of our timescale tables that I'd love to get better compression on. It changes just frequently enough that we didn't split it into columns, but infrequently enough that it could (I think) be compressed quite nicely. Generally timescale has been great for us.
I have been working on another PG extension for timeseries (https://github.com/xataio/deltax) for a few months, and trying to make it 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.
If I understand your question correctly:
For storage, mostly yes. Swinging-door compression went lossy because byte compression on raw floats used to be useless, but with new method like Gorilla you can compress it lossless, so you can keep every sample and still afford it.
But there is another point that this Aveva historian adds filter that filters some points so I don't know how it impact ingest.
But in general now trend is to use normal databases from IT world in OT world to overcome some legacy solutions.
Based on us keeping a replica of our AVEVA data archive in delta tables which are just parquets; and only compressing the parquets in LZ4 and getting a fuck ton of more space savings. I would say yes. Especially when you're onprem with a dedupe storage appliance.
I think swinging door made sense when storage was slow and scrace, but nowadays storage is cheap (even with these prices)
> 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...
There’s an issue tracking TimescaleDB JSONB compression: https://github.com/timescale/timescaledb/issues/2978
Ha, we (my partner at our company) filed this issue 5 years ago. We have a large-ish (but not giant) json blob in one of our timescale tables that I'd love to get better compression on. It changes just frequently enough that we didn't split it into columns, but infrequently enough that it could (I think) be compressed quite nicely. Generally timescale has been great for us.
I have been working on another PG extension for timeseries (https://github.com/xataio/deltax) for a few months, and trying to make it 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.
Old data historians use swinging-door compression algorithms (a lossy compression). Would something like TimescaleDB make this unnecessary for IoT?
https://docs.aveva.com/bundle/pi-server-s-da-admin/page/1022...
If I understand your question correctly: For storage, mostly yes. Swinging-door compression went lossy because byte compression on raw floats used to be useless, but with new method like Gorilla you can compress it lossless, so you can keep every sample and still afford it. But there is another point that this Aveva historian adds filter that filters some points so I don't know how it impact ingest.
But in general now trend is to use normal databases from IT world in OT world to overcome some legacy solutions.
>But in general now trend is to use normal databases from IT world in OT world to overcome some legacy solutions.
Someone tell my company this please ;-;
Based on us keeping a replica of our AVEVA data archive in delta tables which are just parquets; and only compressing the parquets in LZ4 and getting a fuck ton of more space savings. I would say yes. Especially when you're onprem with a dedupe storage appliance.
I think swinging door made sense when storage was slow and scrace, but nowadays storage is cheap (even with these prices)
Up to 100% of articles using "up to" in the title are complete dross.
How do I get the job of benevolent dictator of title moderation?
Okay maybe title is not great but I really achieved this compression on real database where I have devices via MQTT connected.
Gorilla by Facebook had this. Value is stored as delta and time as delta of delta.
It's used in ClickHouse as well. CH supports all known compression algos and they are documented pretty well.
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.
Compression uses a different license than Apache.
In various distros, the timescaledb installed by the package manager doesn't support other licenses than Apache.
This means that there's a high chance your distro doesn't allow stock timescaledb to use compression. You probably will need to install it manually.