I'm genuinely surprised that there isn't column-level shared-dictionary string compression built into SQLite, MySQL/MariaDB or Postgres, like this post is describing.
SQLite has no compression support, MySQL/MariaDB have page-level compression which doesn't work great and I've never seen anyone enable in production, and Postgres has per-value compression which is good for extremely long strings, but useless for short ones.
There are just so many string columns where values and substrings get repeated so much, whether you're storing names, URL's, or just regular text. And I have databases I know would be reduced in size by at least half.
Is it just really really hard to maintain a shared dictionary when constantly adding and deleting values? Is there just no established reference algorithm for it?
It still seems like it would be worth it even if it were something you had to manually set. E.g. wait until your table has 100,000 values, build a dictionary from those, and the dictionary is set in stone and used for the next 10,000,000 rows too unless you rebuild it in the future (which would be an expensive operation).
> Is it just really really hard to maintain a shared dictionary when constantly adding and deleting values? Is there just no established reference algorithm for it?
Enums? Foreign key to a table with (id bigint generated always as identity, text text) ?
> I have databases I know would be reduced in size by at least half.
Most people don't employ these strategies because storage is cheap and compute time is expensive.
Strings in textual index are already compressed, with common prefix compression or other schemes. They are perfectly queryable. Not sure if their compression scheme is for index or data columns.
Global column dictionary has more complexity than normal. Now you are touching more pages than just the index pages and data page. The dictionary entries are sorted, so you need to worry about page expansion and contraction. They sidestep the problems by making it immutable, presumably building it up front by scanning all the data.
Not sure why using FSST is better than using a standard compression algorithm to compress the dictionary entries.
Storing the strings themselves as dictionary IDs is a good idea, as they can be processed quickly with SIMD.
> Not sure why using FSST is better than using a standard compression algorithm to compress the dictionary entries.
I believe the reason is that FSST allows access to individual strings in the compressed corpus, which is required for fast random access. This is more important for OLTP than OLAP, I assume.
More standard compression algorithms, such as zstd, might decompress very fast, but I don't think they allow that
There are some databases that can move an entire column into the index. But that's mostly going to work for schemas where the number of distinct values is <<< rowcount, so that you're effectively interning the rows.
1, complicates and slows down update, which is typically more important in OLTP than OLAP
2, is generally bad for high cardinality columns, which requires tracking cardinality to make decisions, which further complicates things.
lastly, additional operational complexity (like the table maintenance system you described in last paragraph) could reduce system reliability, and they might decide it's not worth the price or against their philosophy.
But string interning is what they're doing, isn't it?
> Dictionary compression is a well-known and widely applied technique. The basic idea is that you store all the unique input values within a dictionary, and then you compress the input by substituting the input values with smaller fixed-size integer keys that act as offsets into the dictionary. Building a CedarDB dictionary on our input data and compressing the data would look like this:
That's string interning!!
Is interning just too old a concept now and it has to be rediscovered/reinvented and renamed?
String interning only stores whole strings, dictionary compression stores substrings. Essentially string interning is a trivial subset of dictionary compression (because you don’t need a substringing scheme, which is the hard part).
Doesn't interning usually refer to when you only consider identical copies, as opposed to dictionary compression where you allow for concatenations? E.g.
Interning:
1: "foo"
2: "bar"
my_string = "foo" // stored as ref->1
my_other_string = "foobarbaz" // not found & too long to get interned, stored as "foobarbaz"
Dictionary compression:
1: "foo"
2: "bar"
my_string = "foo" // stored as ref->1
my_other_string = "foobarbaz" // stored as ref->1,ref->2,"baz" (or ref->1,ref->2,ref->3 and "baz" is added to the dict)
Only if you explicitly intern the string. Interning can be expensive because
- the runtime has to check whether the string already is interned.
- you’re permanently creating the string, so if it isn’t reused later, that means your memory usage needlessly goes up.
Both get more expensive the more strings you intern.
I think interning is a hack that very rarely, if ever should be used. It likely won’t work well for large strings, as these tend to be unique, and use cases where it helps for shirt strings often are better handled by using enums or symbols, or by using a custom set of strings. If you do the latter, you have more control over Emory usage; you can do such things such as removing the least recently used strings, or ditching the entire cache when you’re done needing it (prime example: parsing a large XML file with many repeated nodes)
The compression algorithm is very similar to a greedy subword tokenizer, which is used in BERT and other older language models, but has become less popular in favor of BPE.
I've implemented a similar system based on the original 2020 paper,
but we applied it to the text log to try to "extract" similar features
from free-form text. It looked promising and even supported full
regex search, but the work was ultimately abandoned when we got
acquired.
I was evaluating it recently but it's not FOSS, so buyer beware. I'm totally fine with commercialization, but I hesitate to build on top of data stores with no escape hatches or maintenance plans–especially when they're venture backed. It is self-hostable, but not OSS.
It's probably going to be acquired. The last effort to commercialize the TUM (Technical University of Munich) database group's work was acquired by Snowflake and disappeared into that stack.
CedarDB is the commercialization of Umbra, the TUM group's in-memory database lead by professor Thomas Neumann. Umbra is a successor to HyPer, so this is the third generation of the system Neumann came up with.
Umbra/CedarDB isn't a completely new way of doing database stuff, but basically a combination of several things that rearchitect the query engine from the ground up for modern systems: A query compiler that generates native code, a buffer pool manager optimized for multi core, push-based DAG execution that divides work into batches ("morsels"), and in-memory Adaptive Radix Tries (never used in a database before, I think).
It also has an advanced query planner that embraces the latest theoretical advances in query optimization, especially some techniques to unnest complex multi-join query plans, especially with queries that have a ton of joins. The TUM group has published some great papers on this.
> It also has an advanced query planner that embraces the latest theoretical advances in query optimization, especially some techniques to unnest complex multi-join query plans, especially with queries that have a ton of joins. The TUM group has published some great papers on this.
I always wondered how good these planners are in practice. The Neumann/Moerkotte papers are top notch (I've implemented several of them myself), but a planner is much more than its theoretical capabilities; you need so much tweaking and tuning to make anything work well, especially in the cost model. Does anyone have any Umbra experience and can say how well it works for things that are not DBT-3?
Umbra is not an in-memory database (Hyper was). TUM gave up on the feasibility of in-memory databases several years ago (when the price of RAM relative to storage stopped falling).
Yeah I think the way Umbra was pitched when I watched the talks and read the paper was as more as "hybrid" in the sense that it aimed for something close to in-memory performance while optimizing the page-in/page-out performance profile.
The part of Umbra I found interesting was the buffer pool, so that's where focused most of my attention when reading though.
This isn't strictly correct: you probably mean wrt compressed size. Compression is a tradeoff between size reduction and compression and decompression speed. So while things like Bellard's tz_zip (https://bellard.org/ts_zip/) or nncp compress really well they are extremely slow compared to say zstd or the much faster compression scheme in the article. It's a totally different class of codec.
SQLite has no compression support, MySQL/MariaDB have page-level compression which doesn't work great and I've never seen anyone enable in production, and Postgres has per-value compression which is good for extremely long strings, but useless for short ones.
There are just so many string columns where values and substrings get repeated so much, whether you're storing names, URL's, or just regular text. And I have databases I know would be reduced in size by at least half.
Is it just really really hard to maintain a shared dictionary when constantly adding and deleting values? Is there just no established reference algorithm for it?
It still seems like it would be worth it even if it were something you had to manually set. E.g. wait until your table has 100,000 values, build a dictionary from those, and the dictionary is set in stone and used for the next 10,000,000 rows too unless you rebuild it in the future (which would be an expensive operation).
reply