Regarding hash joins, the SQLite documentation mentions the absence of real hash tables [0]
SQLite constructs a transient index instead of a hash table in this instance
because it already has a robust and high performance B-Tree implementation at
hand, whereas a hash-table would need to be added. Adding a separate hash table
implementation to handle this one case would increase the size of the library
(which is designed for use on low-memory embedded devices) for minimal
performance gain.
It's already linked in the paper, but here's the link to the code used in the paper [1]
The paper mentions implementing Bloom filters for analytical queries an explains how they're used. I wonder if this is related to the query planner enhancements that landed on SQLite 3.38.0 [2]
Use a Bloom filter to speed up large analytic queries.
That's correct, the optimizations from this paper became available in SQLite version 3.38.0.
As we were writing the paper, we did consider implementing hash joins in SQLite. However, we ultimately went with the Bloom filter methods because they resulted in large performance gains for minimal added complexity (2 virtual instructions, a simple data structure, and a small change to the query planner). Hash joins may indeed provide some additional performance gains, but the question (as noted above) is whether they are worth the added complexity.
TFA appears to be about adapting SQLite for OLAP workloads. I do not understand the rationale. Why try to adapt a row-based storage system for OLAP? Why not just use a column store?
SQLite is significantly better at OLTP and being a blob strorage than DuckDB, and it doesn't want to sacrifice those advantages and compatibility if OLAP performance can be improved independently. In my experience for diverse workloads it is more practical to start with a row-based structure and incrementally transform it into a column-based one. Indeed in the paper there is a suggested approach that trades space for improved OLAP performance.
It is certainly possible to have a single system that can effectively process high volumes of OLTP traffic while at the same time performing OLAP operations. While there are systems that are designed to do one or the other type of operation well, very few are able to do both. https://www.youtube.com/watch?v=F6-O9v4mrCc
It seems like one idea in there is to store it both ways automatically (the HE variant)! That might be better then manually continually copying between your row store and your column store.
Great discussion here. As one of the co-authors of the paper, here is some additional information.
If you need both transactions and OLAP in the same system, the prevalent way to deliver high performance on this (HTAP) workload is to make two copies of the data. This is what we did in the SQLite3/HE work (paper: https://www.cidrdb.org/cidr2022/papers/p56-prammer.pdf; talk: https://www.youtube.com/watch?v=c9bQyzm6JRU). That was quite clunky. This two copy approach not only wasted storage but makes the code complicated, and it would be very hard to maintain over time (we did not want to fork the SQLite code -- that is not nice).
So, we approached it in a different way and started to look for how we could get higher performance on OLAP queries working as closely with SQLite's native query processing and storage framework.
There is a lot of room for improvement here, and getting high OLAP and transaction performance in a single-copy database system is IMO a holy grail that many in the community are working on.
BTW - the SQLite team, namely Dr. Hipp (that is a cool name), Lawrence and Dan are amazing to work with. As an academic, I very much enjoyed how deeply academic they are in their thinking. No surprise that they have built an amazing data platform (I call it a data platform as it is much more than a database system, as it has many hooks for extensibility).
This is basically the exact use case SQLite was designed for; PostgreSQL is a marvel, and at the end of the day presents a much more robust RDBMS, but it's never going to beat SQLite at the thing SQLite was designed for.
I am 99% sure SQLite is going to win unless you actually care about data durability at power loss time. Even if you do, I feel I could defeat Postgres on equal terms if you permit me access to certain ring-buffer-style, micro-batching, inter-thread communication primitives.
Sqlite is not great at dealing with a gigantic wall of concurrent requests out of the box, but using a little bit of innovation in front of SQLite can solve this problem quite well. The key is resolve the write contention outside of the lock that is baked into the SQLite connection. Writing batches to SQLite on a single connection with WAL turned on and Sync set to normal is pretty much like operating at line speed with your IO subsystem.
"An SQLite database is highly resistant to corruption. If an application crash, or an operating-system crash, or even a power failure occurs in the middle of a transaction, the partially written transaction should be automatically rolled back the next time the database file is accessed. The recovery process is fully automatic and does not require any action on the part of the user or the application."
"Crash testing seeks to demonstrate that an SQLite database will not go corrupt if the application or operating system crashes or if there is a power failure in the middle of a database update. A separate white-paper titled Atomic Commit in SQLite describes the defensive measure SQLite takes to prevent database corruption following a crash. Crash tests strive to verify that those defensive measures are working correctly.
It is impractical to do crash testing using real power failures, of course, and so crash testing is done in simulation. An alternative Virtual File System is inserted that allows the test harness to simulate the state of the database file following a crash."
By default WAL mode can rollback committed transactions in cases of power failure, but you can do `PRAGMA synchronous = FULL` to trade speed for durability.
If you can have one "database" thread and 63 "worker" threads, send messages back and forth, and don't hold open transactions, this would probably work with sqlite. Aka treat sqlite like redis.
so in your example the database thread is the Redis thread and the worker thread are your http server thread I assume.
This is a good analogy, but there are still lot of wire heavy scenario a real database like postgresql or mysql will have better throughput than redis.
Redis has the same limitation (only one transaction at a time) and is used a lot for webapps. It solves this by requiring full transactions up front. The ideal case for sqlite for performance is to have only a single process/thread directly interacting with the database and having other process/threads send messages to and from the database process.
As long as you turn it into a throughput race instead of a latency race, PostgreSQL can definitely win. SQLite has a primitive query builder and a limited selection of query execution steps to choose from. For instance, all joins in SQLite are inner loop joins. It can't do hash or merge joins. It can't do GIN or columnstore indexes. If a query needs those things, PostgreSQL can provide them and can beat SQLite.
out of interest, what columnstore indexes are available to postgres? Would be happy to find out that I'm missing something.
I know citus can provide columnar tables but I can't find columnar indexes for regular row-based tables in their docs. (use case of keeping an OLTP table but wanting to speed up a tiny subset of queries)
Closest thing I could find was Swarm64 for columnar indexes but it doesn't seem to be available anymore.
> just from the fact that there are less layers of code to be worked through
This is not an invariant. I've seen be true, and I've seen it be false. Sometimes that extra code is just cruft yes. Other times though it is worth it to set up your data (or whatever) to take advantage of mechanical sympathies in hot paths, or filter the data before the expensive processing step, etc.
I'm not talking about extra code, I'm talking about _layers_ of code. With PostgreSQL you're still sending data over TCP/IP or a UNIX socket, and are copying things around in memory. Compare that to SQLite that runs in the memory space of the program, thus no need for copying and socket traffic. There's just less middlemen (middlepersons?) with SQLite that are unavoidable with PostgreSQL. So less layers = less interpreting/serialization/deserialization/copying/... = higher performance. I will even argue that even if the SQLite query engine is slightly less efficient than PostgreSQL, you're still winning because of less memory copying going around.
> less interpreting/serialization/deserialization/copying/... = higher performance
Unfortunately for many database workloads you are overestimating the relative cost of this factor.
> even if the SQLite query engine is slightly less efficient than PostgreSQL
And this is absurd - the postgresql query engine isn't just "slightly" more efficient. It is tremendously more sophisticated. People using a SQL datastore as a glorified key-value store are not going to notice - which seems to be a large percentage of the sqlite install base.
It's not really a fair comparison.
With SQLite, though, you could reasonably just skip doing fancy joins and do everything in tiny queries in tight loops because SQLite is literally embedded in your app’s code. You can be careless with SQLite in ways you cannot with a monolithic database server because of that reason. I still agree there are use cases where a centralized database is better, but SQLite is a strange beast that needs a special diet to perform best.
Sometimes. This is only universally true if your datasets are tiny and your access patterns are simple. Moving a shitty suboptimal O(n^2) or worse algorithm locally when something linear or better is possible is going to hurt no matter where the DB is.
> but SQLite is a strange beast that needs a special diet to perform best.
I don’t see what is strange about it - for large datasets it’s the same complexity issues as anywhere else.
Not sure specifically what your comment is trying to add, since I acknowledged the type of use case SQLite excels in - those where roundtripping are a dominating cost and “k-v” stores, ie simple queries. My entire point is that those are a common but still niche use case.
Then it depends on the usage. You'd likely need to run with synchronous mode on, and even on WAL, multiple separate write transactions is a issue. If you don't have many writes or buffer them into not many transactions, SQLite is the most performant.
I think the (unsatisfying) answer is "it depends". There's a huge amount of diversity in database workloads, even among the workloads served by SQLite as we mention in the paper.
For read-mostly to read-only OLTP workloads, read latency is the most important factor, so I predict SQLite would have an edge over PostgreSQL due to SQLite's lower complexity and lack of interprocess communication.
For write-heavy OLTP workloads, coordinating concurrent writes becomes important, so I predict PostgreSQL would provide higher throughput than SQLite because PostgreSQL allows more concurrency.
For OLAP workloads, it's less clear. As a client-server database system, PostgreSQL can afford to be more aggressive with memory usage and parallelism. In contrast, SQLite uses memory sparingly and provides minimal intra-query parallelism. If you pressed me to make a prediction, I'd probably say SQLite would generally win for smaller databases. PostgreSQL might be faster for some workloads on larger databases. However, these are just guesses and the only way to be sure is to actually run some benchmarks.
Is it the SQL dialect there lacking or is it the built-in functions?
I agree that SQLite default functionality is very thin compared to PostgreSQL - especially with respect to things like date manipulation - but you can extend it with more SQL functions (and table-valued functions) very easily.
I like SQLite (qualifying not for you, simonw, but for others). But I hate that I can't be lazy by using arrays in SQLite... because they don't exist. group_concat is a poor approximation.
Also, I genuinely dislike how loose SQLite is with allowed syntax. Probably it's preference. But even interactively I prefer to know immediately that I messed up a query. SQLite is so forgiving I've often wasted time trying to understand why my results are nonsense (because I typoed in the query and SQLite didn't fail the query).
But I also strongly dislike Python for that reason and I know where you stand there. Maybe SQLite/PostgreSQL is similar to the dynamic/static language preference divide.
I'm increasingly finding myself using SQLite's JSON features for array stuff - they're surprisingly convenient once you get into the habit of using them. A couple of recent examples:
Adding user-defined functions to SQLite is not difficult, and the mechanism is quite flexible. You can create extensions and load them when you create the SQLite connection to have the functions available in queries. I wrote a blog post explaining how to do that using Rust, and the example is precisely a `regex_extract` function [0].
If you need them, you also have a "stdlib" implemented for Go [1] and a pretty extensive collection of extensions [2]
Probably also worth noting: you don't need to build (many kinds of) extensions as C-compatible code and separate .so files that you load.
SQLite is an in-process database. You can give it a callback func to execute. So your regex-extract can literally just call a function in your code: https://sqlite.org/appfunc.html
edit: Python's stdlib documentation concisely shows how easy this can be: https://docs.python.org/3/library/sqlite3.html#sqlite3.Conne... Basically every SQLite library should have something similar. This extreme ease of extending is a big part of why SQLite has so little built-in.
Ha ha .. that is what the conference requires. Turns out that there is research that shows that when you are reading paper printed on paper this 2-column format is good for readability and not wasting paper. Conferences still insist on this format even though most people print papers.
Now the good news is that these days, conferences have an accompanying video associated with the paper, and that may be a good place to start for many. That video will be published on the conference website (https://vldb.org/2022/) in about a week.
I'd like to see that. I also think the single write situation is not great for web applications, but I don't see an easy way around it without sacrificing things like consistency
See: LMAX Disruptor and friends. The magic spell that serializes many threads of events into one without relying on hard contention. You can even skip the hot busy waiting if you aren't trying to arbitrage the stock market.
The way I would do it is a MPSC setup wherein the single consumer holds an exclusive connection to the SQLite database and writes out transactions in terms of the batch size. Basically: BEGIN -> iterate & process event batch -> END. This is very very fast due to how the CPU works at hardware level. It's also a good place to insert stuff like [a]synchronous replication logic.
Completion is handled with busy/yield waiting on a status flag attached to the original event instance. You'd typically do a thing where all flags are acknowledged at the batch grain (i.e. after you committed the txn). This has some overhead, but the throughput & latency figures are really hard to argue with. It also has very compelling characteristics at the extremes in terms of system load. The harder you push, the faster it goes.
my ipad won’t let me search through the PDF, but i couldn’t find where “SSB” was defined, if anywhere. i did not see it defined in the first paragraph, which is where it is first used.
everyone: not all of your readers are domain experts. omissions like this are infuriating.
Just fyi: if you’re viewing the PDF in Safari on your iPad, you can search by typing into Safari’s Location Bar and then choosing “Find ‘xyz’” from the popup that appears.
"While it continues to be the most widely used database engine in the world"
It realy depends what do you mean by that, yes it's shipping in every phones and browser, but I don't consider that as a database. Is the windows registry a database?
Oracle, MySQL, PG, MSSQL are the most widly used DB in the world, the web runs on those not SQLite.
there are far, far more sqlite instances than Windows Registry instances in the world.
"SQLite is likely used more than all other database engines combined. Billions and billions of copies of SQLite exist in the wild. [...] Since SQLite is used extensively in every smartphone, and there are more than 4.0 billion (4.0e9) smartphones in active use, each holding hundreds of SQLite database files, it is seems likely that there are over one trillion (1e12) SQLite databases in active use."
> SQLite is pretty bad when it comes to transactions per second. SQLite even owns up to it in the FAQ: "it will only do a few dozen transactions per second."
That is an extremely poor quote taken way out of context.
The full quote is:
FAQ: "[Question] INSERT is really slow - I can only do few dozen INSERTs per second. [Answer] Actually, SQLite will easily do 50,000 or more INSERT statements per second on an average desktop computer. But it will only do a few dozen transactions per second. Transaction speed is limited by the rotational speed of your disk drive. A transaction normally requires two complete rotations of the disk platter, which on a 7200RPM disk drive limits you to about 60 transactions per second."
Yeah my GP got me confused. I remember doing 40k inserts/s in a trading strategy backtesting program with Go and SQLite. Reads were on the same magnitude, I want to say around 90k/s. My bottleneck was CPU.
Thanks for the clarification. It's true that transaction latency is limited by the write speed of the storage medium. However, an "average desktop computer" these days has an SSD that can support tens of thousands of transactions per second, depending on the workload.
What is your point? If I need transactions, not just bulk loading inserts, then SQLite isn't the bees knees. PG can handle at least an order of magnitude more transactions per second on the same hardware.
Please quote the entire statement. And stop the needless "even owns up to it" FUD.
> Actually, SQLite will easily do 50,000 or more INSERT statements per second on an average desktop computer. But it will only do a few dozen transactions per second. Transaction speed is limited by the rotational speed of your disk drive. A transaction normally requires two complete rotations of the disk platter, which on a 7200RPM disk drive limits you to about 60 transactions per second.
