Better JIT for Postgres
github.com
github.com
I don't know anything here, but this seems like a good case for ahead of time compilation? Or at least caching your JIT results? I can image much of the time, you are getting more or less the same query again and again?
Some years ago we ported some code from querying out the data and tallying in Python (how many are in each bucket) to using SQL to do that. It didn't speed up the execution. I was surprised by that, but I guess the Postgres interpreter is roughly the same speed as Python, which when you think about it perhaps isn't that surprising.
But Python is truly general purpose while the core query stuff in SQL is really specialized (we were not using stored procedures). So if Pypy can get 5x speedup, it seems to me that it should be possible to get the same kind of speed up in Postgres. I guess it needs funding and someone as smart as the Pypy people.
MS SQL still has prepared statements and they really haven't been used in 20 years since it gained the ability to cache plans based on statement text.
And then the issue is not dissimilar to Postgres’s planner issues.
The problem though was that it had a single shared pool for all queries and it could only run a query if it was in the pool, which is how out DB machine would max out at 50% CPU and bandwidth. We had made some mistakes in our search code that I told the engineer not to make.
- not all databases need migrations (or migrations without downtime)
- alternatively, ship the migrations as part of the binary
Adhoc modifications would still be more difficult but tbh that’s not necessarily a bug
> Figure 2 shows the experimental results, and GenDB outperforms all baselines on every query in both benchmarks. On TPC-H, GenDB achieves a total execution time of 214 ms across five representative queries.
> This result is 2.8× faster than DuckDB (594 ms) and Umbra (590 ms), which are the two fastest baselines, and 11.2× faster than ClickHouse.
> On SEC-EDGAR, GenDB achieves 328 ms, which is 5.0× faster than DuckDB and 3.9× faster than Umbra.
> The performance gap increases with query complexity. For example, on TPC-H Q9, which is a five-way join with a LIKE filter, GenDB completes in 38 ms, which is 6.1× faster than DuckDB. GenDB uses iterative optimization with early stopping criteria.
> On TPC-H, Q6 reaches a near-optimal time of 17 ms at iteration 0 with zone-map pruning and a branchless scan, and does not require further optimization. In contrast, Q18 starts at 12,147 ms and decreases to 74 ms by iteration 1, which is a 163× improvement. This gain comes from replacing a cache-thrashing hash aggregation with an index-aware sequential scan.
> On SEC-EDGAR, Q4 decreases from 1,410 ms to 106 ms over three iterations, which is a 13.3× improvement, and Q6 decreases from 1,121 ms to 88 ms over four iterations, which is a 12.7× improvement. In Q6, the optimizer gradually fuses scan, compact, and merge operations into a single OpenMP parallel region, which removes three thread-spawn overheads. By iteration 1, GenDB already outperforms all baselines
And knowing typical LLM latency, it's outside of the realm of OLTP and probably even OLAP. You can't wait tens of seconds to minutes until LLM generates you some optimal code that you then compile and execute.
The catch is obviously that the development is way too expensive and that it takes a lot of technical capability which isn't really all that common. The novelty which this paper presents is that these two barriers might have come to an end - we can use LLMs and agents to build custom database engines for ourselves™ and our™ specific workloads, very quickly and for a tiny fraction of development price.
Most other DB's cache query plans including jitted code so they are basically precompiled from one request to the next with the same statement.
I was actually confused by this submission as it puts so much of an emphasis on initial compilation time, when every DB (apparently except for pgsql) caches that result and shares it/reuses it until invalidation. Invalidation can occur for a wide variety of reasons (data composition changing, age, etc), but still the idea of redoing it on every query, where most DBs see the same queries endlessly, is insane.
To make code shareable between processes takes effort and will have tradeoff in performance since it is not specialized to the process.
If the query plan where at least serializable which is more like a AST then at least that part could be reused and then maybe have jitted code in each processes cached in memory that the plan can reference by some key.
DB's like MSSQL avoid the problem because they run a single OS process with multiple threads instead. This is also why it can handle more connections easily since each connection is not a whole process.
The current PG query plan and jit is designed around just being in memory in a single process, this would need to be extracted into something not process specific and shared between all processes. The plan itself is just a bunch of C structs I believe.
"This obvious drawback of the current software architecture motivates our work: sharing JIT code caches across applications. During the exploration of this idea, we have encountered several challenges. First of all, most JIT compilers leverage both runtime context and profile information to generate optimized code. The compiled code may be embedded with runtime-specific pointers, simplified through unique class-hierarchy analysis, or inlined recursively. Each of these "improve- ments" can decrease the shareability of JIT compiled code."
Anythings doable here with enough dev time. Would be nice if PG could just serialize the query plan itself maybe just as an SO along with non-process specific executable code that then has to be dynamically linked again in other processes.
https://www.postgresql.org/docs/current/parallel-query.html
"PostgreSQL can devise query plans that can leverage multiple CPUs in order to answer queries faster."
If process based then they can send small parts of plan across processes.
Plans for prepared statements are cached though.
Other databases like MSSQL have prepared statements but they are rarely used now days since plan caching based on query text was introduced decades ago.
The problems related to PostgreSQL are pretty much all described here. It's very difficult to do low-latency queries if you cannot cache the compiled code and do it over and over again. And once your JIT is slow you need a logic to decide whether to interpret or compile.
I think it would be the best to start interpreting the query and start compilation in another thread, and once the compilation is finished and interpreter still running, stop the interpreter and run the JIT compiled code. This would give you the best latency, because there would be no waiting for JIT compiler.
This is not too difficult, it just requires a different execution style. Salesforce's Hyper for example very heavily relies on JIT compilation, as does Umbra [1], which some people regard as one of the fastest databases right now. Umbra doesn't cache any IR or compiled code and still has an extremely low start-up latency; an interpreter exists but is practically never used.
Postgres is very robust and very powerful, but simply not designed for fast execution of queries.
Disclosure: I work in the group that develops Umbra.
The problem will always be queries where the compilation is orders of magnitude more expensive than the query itself. I can imagine indexed lookup of 1 or few entries, etc... Accessing indexed entries like these are very well optimized by SQL query engines and possibly make no sense JIT optimizing.
I think AsmJit's strength is completeness of its backends as you can emit nice SIMD code with it (like AVX-512). But the performance could be better of course, and that's possible - making it 2x faster would be possible.
If this function is optimized, or switched to some other implementation when there is tens of thousands of virtual registers, you would get orders of magnitude faster compilation.
But realistically, which query requires tens of megabytes of machine code? These are pathological cases. For example we are talking about 25ms when it comes to a single function having 1MB of machine code, and sub-ms time when you generate tens of KB of machine code.
So from my perspective the ability to generate SIMD code that the CPU would execute fast in inner loops is much more valuable than anything else. Any workload, which is CPU-bound just deserves this. The question is how much the CPU bound the workload is. I would imagine databases like postgres would be more memory-bound if you are processing huge rows and accessing only a very tiny part of each row - that's why columnar databases are so popular, but of course they have different problems.
I worked on one project, which tried to deal with this by using buckets and hashing in a way that there would be 16 buckets, and each column would get into one of these, to make the columns closer to each other, so the query engine needs to load only buckets used in the query. But we are talking about gigabytes of RAW throughput per core in this case.
Definitely good to know though. When it comes to low-latency compilation my personal goal is to make it even faster when generating small functions.
But... I consider SLJIT to be for a different use-case than AsmJit. It's more portable, but its scope is much more limited.
This technique is known as a "tiered JIT". It's how production virtual machines operate for high-level languages like JavaScript.
There can be many tiers, like an interpreter, baseline compiler, optimizing compiler, etc. The runtime switches into the faster tier once it becomes ready.
More info for the interested:
So there is not much to gain from JITing the query plan execution only.
JITing begins to make more sense, when the individual query plan steps (join, filter, ...) themselves be specialized/recompiled/improved/merged by knowing the context of the query plan.
[1]: https://pkg.go.dev/github.com/jackc/pgx/v5#hdr-Prepared_Stat...
[2]: https://www.psycopg.org/psycopg3/docs/advanced/prepare.html
Ad-hoc one off queries usually can accept higher initial up-front compile cost because the main results usually take much longer anyway, vs worrying about an extra 100ms of compile.
Maybe it was too strong to say its not a concern at all, but nothing like PG where every single request needs to replan and potentially jit unless the client manually prepares and keeps the connection open.
Was common guidance back in the day to use stored procedures for all application access code because they where cached in MSSQL (which PG doesn't even do). Then around 2000 it started caching based on statement text and that became much less important.
You would only used prepared statements if doing a bunch of inserts in a loop or something and it has a very small benefit now days only because its not sending the same text over the network over and over and hashing to lookup plan.
They also do things like auto parameterization if the statement doesn't have them and parameter sniffing to make multiple different plans based on different values where it makes sense.
https://learn.microsoft.com/en-us/sql/relational-databases/q...
You can also get this, add HINTs to control this behavior if you don't like it or its causing a problem in production, crazy I know.
https://learn.microsoft.com/en-us/sql/t-sql/queries/hints-tr...
PG is extremely primitive compared to these other systems in this area, and it has to be since it doesn't cache anything unless specifically instructed to for a single connection.
I never said it was simple, in fact I said how primitive PG is compared to the "big boys" because they put huge effort into making their systems fast back in the TPS wars of the early 2000's on much slower hardware.
>Prepared statements != cached execution plans
Thats exactly what a prepared statement is:
So narrow its enabled by default for all statements from the "big boy" commercial RDBMS's...
https://www.ibm.com/docs/en/i/7.4.0?topic=overview-plan-cach...
https://docs.oracle.com/en/database/oracle/oracle-database/1...
https://learn.microsoft.com/en-us/sql/relational-databases/p...
https://help.sap.com/docs/SAP_HANA_PLATFORM/6b94445c94ae495c...
>Postgres cached plans exist for the same reason.
Postgresql doesn't cache plans unless the client explicitly sends commands to do so. Applications cannot take advantage of this unless they keep connections open and reuse them in a pool and they must mange this themselves. The plan has to be planned for every separate connection/process rather than a single cached planed increasing server memory costs which are plan cache size X number of connections.
It has no "reason" to cache plans the client must do this using its "reasons".
>If you're claiming Oracle and MSSQL do _much_ better in this area - that's what I call unsubstantiated.
You are making all sorts of claims without nary a link to back it up. Are you suggestion PG does better than MSSQL, Oracle and DB2 in planning while be constrained to replan on every single statement? The PG planner is specifically kept simple so that it is fast at its job, not thorough or it would adversely effect execution time more than it already does, this is well documented and always a concern when new features are proposed for it.
>From what you write further it's pretty clear you don't have a lot of understanding what happens under the hood.
Sticks and stones, is that all you have how about something substantial.
> And no, prepared statements are not what you read in Wikipedia. Not in all databases anyway.
Ok Mr. Unsubstantiated are we talking about PG or not? What does one use prepared statements for in PG hmmm, you know the thing you call the PG plan cache? How about something besides your claim that prepared statements are not in fact plan caches? Are you talking about completely different DB systems? How about you substantiate that?
and then
https://www.postgresql.org/docs/current/sql-prepare.html
Read carefully about "plan_cache_mode" and how it works (and its default settings). Sorry, that's my last message in this thread, and I'm still here just for educational purposes, because what you're talking about is in fact a common misconception. If you read it carefully, you'll see that generic plans do not require any "explicit commands", Postgres executes a query 5 times in custom mode, then tries a generic one, if it worked (not much worse than an average of 5 custom plans), the plan is cached. You can turn it off though. And I'd recommend to turn it off for most cases, because it's a pretty bad heuristics. Nevertheless, for some (pretty narrow set of) cases it's useful.
So, Mr Big Boy, now we can get to what a prepared statement in Postgres is. Prepared statements are cached in a session, but if that statement was cached in custom mode, it won't contain a plan. When Postgres receives a prepared statement in custom mode, it will just skip parsing, that's it. The query will still be planned, because custom plans rely on input parameters. If we run it in generic mode, then the plan is cached.
Seems like it's not great at detecting this in all cases[1]. That said, I do note that was reproduced on PG16, perhaps they've made improvements since, given the documentation explicitly mentions what you said.
[1]: https://www.michal-drozd.com/en/blog/postgresql-prepared-sta...
This behavior is often a source of random latency spikes, when your queries suddenly start misbehaving, and then suddenly stop doing it. If you don't have auto_explain on, it will look like mysterious glitches in production.
The few cases when they are useful are very simple ones, like single table selects by index. They are already fast, and with generic plans you can cut planning time completely. Which is kinda...not much. There are more complicated cases where they are useful, involving Postgres forks like AWS Aurora, which has query plan management subsystem, allowing to store plans directly. Then you can cut planning time for them. But that's a completely different story.
This controls whether prepared statements even use a cached plan at all. Other database can do this with hints and they can skip parsing by using stored procedures which are basically globally named prepared statements that the client can call without preparing a temporary one or they can do prepared but again this is typically a waste of time because parsing enough to match existing plans is fast (soft vs hard parse in Oracle speak). They have many more options with more powerful caching abilities that all clients can share across sessions.
The only time PG "automatically" caches the plan is when it implicitly prepares the plan within a PL/pgsql statement like doing a insert loop inside a function, its still is only for the current session. This is just part of the planning process in other databases that cache everything all the time globally.
You don't seem to understand that most other commercial "big-boy" RDBMS cache plans across sessions and that nothing has to be done for them to reuse between completely different connections with differing parameters and can still have specialized versions based on these parameters values vs a single generic plan.
At least now you admit prepared statements are in-fact a plan cache, contradicting your other statements, and seem to make a gotcha out of an option an option to disable that cache.
You can see various discussions on pg-hackers, here is one where the submitter confirms everything I have said and attempted to add the auto part but not tackle the much harder sharing between sessions part and was shot down, I don't believe much has change in PG around plan caching since this post and even has a guy that worked on DB2 talking about how they did it: https://www.postgresql.org/message-id/flat/8e76d8fc-8b8c-14b...
Please substantiate this, again all other major commercial RDBMS's do this and have invested a lot of effort and money into these systems, they would not do something that has marginal value.
Again I went through the era of needing to manually prepare queries in client code when it was the only choice as it is now in PG. It was not a marginal improvement when automatic global caching became available, it was objectively measurable via industry standard benchmarks.
You can also find other post complaining about prepared statement cache memory usage especially when libs and pooler auto prepare, the cache is repeated for every connection, 100 connections equals 100X cache size. Another advantage of a shared cache, this is obvious.
I will leave you with a quote from Bruce Momjian, you know one of the founding members of the PG dev team, in the thread I linked that you didn't seem to read just like the other links I gave you:
"I think everyone agrees on the Desirability of the feature, but the Design is the tricky part."
>The main issue you still do not understand is for different parameters you _need_ different plans, and caching doesn't help with that.
You still don't seem to be grasping what other more advanced systems do here and again don't seem to be reading any of the existing literature I am giving you. These systems will make different plans if they detect its necessary, they have MULTIPLE cached plans of the same statement and you can examine their caches and see stats on their usage.
These systems also have hints that let you disable, force a single generic, tell it how you want to evaluate specific parameters for unknown values, specific hard coded values etc. if you want to override their default behavior that uses statistics and heuristic to make a determination of which plan to use.
Please I beg you read what a modern commercial DB can do here and stop saying it doesn't help or can't be done, here is a direct link: https://learn.microsoft.com/en-us/sql/relational-databases/p...
>And you keep conflating "prapared" statements with plan caching.
Again we are talking about PG and the only way PG caches a plan is using prepared statements, in PG prepared statements and plan caching are the same thing, there is no other choice.
From your own link trying to gotcha me on PG plan caching config, first sentence of plan_cache_mode: "Prepared statements (either explicitly prepared or implicitly generated, for example by PL/pgSQL) can be executed using custom or generic plans."
The only other things a prepared statement does is skip parsing, which is another part of caching, and reduce network traffic from client to server. These things can be done with stored procedures in systems that have global caches and are shared across all connections and these systems still support the very rare situation of using a prepared statement, its almost vestigial now days.
Here is Microsoft Guidance on prepared statements in MSSQL now days:
"In SQL Server, the prepare/execute model has no significant performance advantage over direct execution, because of the way SQL Server reuses execution plans. SQL Server has efficient algorithms for matching current Transact-SQL statements with execution plans that are generated for prior executions of the same Transact-SQL statement. If an application executes a Transact-SQL statement with parameter markers multiple times, SQL Server will reuse the execution plan from the first execution for the second and subsequent executions (unless the plan ages from the plan cache)."
https://learn.microsoft.com/en-us/sql/relational-databases/q...
Yeah not a gotcha at all mr teacher. I think you should stop posting low effort responses and examine your own opportunities for education that may have been missed here.
Lets get this straight prepared statements should not be conflated with caching, yet the only way to cache a plan and avoid a full parse is to use a prepared statement and it is by far the biggest reason to use it and why many poolers and libraries try to prepare statements.
Do you realize how ridiculous this is, here is PG's own docs on the purpose of preparing:
"Prepared statements potentially have the largest performance advantage when a single session is being used to execute a large number of similar statements. The performance difference will be particularly significant if the statements are complex to plan or rewrite"
"Although the main point of a prepared statement is to avoid repeated parse analysis and planning of the statement, PostgreSQL will force re-analysis and re-planning of the statement before using it whenever database objects used in the statement have undergone definitional (DDL) changes or their planner statistics have been updated since the previous use of the prepared statement."
The MAIN POINT of preparing is what I am conflating with it, yes...
If PG cached plans automatically and globally then settings like constraint_exclusion and enable_partition_pruning would not need to exist or at least be on by default because the added overhead of the optimizations during planning would be meaningless.
Seriously this whole thread is Brandolini's law in action you obviously can't articulate how PG is better because it does not have a global plan cache and act like I don't know how PG works? Get real buddy.
Are you going to post another couple sentences with no content or are you done here?
> A prepared statement can be executed with either a generic plan or a custom plan. A generic plan is the same across all executions, while a custom plan is generated for a specific execution using the parameter values given in that call.
here https://www.postgresql.org/docs/current/sql-prepare.html
You're also mixing up parsing and planning for some reason. Query parsing costs like 1/100 of planning, it's not nothing, but pretty close to it.
Even though you're just a rude nobody, it still may be useful for others, who may read this stupid conversation...
What is the purpose of a prepared statement without a plan cache? I thought parsing was a non issue? All thats left is a little extra network traffic savings.
I will for a second time quote the PG documentation that you linked btw of what the MAIN POINT of a prepared statement is according to the maintainers, I am not sure why I have to repeat this again:
"Although the main point of a prepared statement is to avoid repeated parse analysis and planning of the statement, PostgreSQL will force re-analysis and re-planning of the statement before using it whenever database objects used in the statement have undergone definitional (DDL) changes or their planner statistics have been updated since the previous use of the prepared statement.”
I am not sure what point you are trying to make other than worming your way out of your previous statements. Prepared statements are in fact plan caches and it is their MAIN purpose according the PG’s own documentation, you haven't given any other purpose for their existence, I gave the other two, one of which you dismissed, the third is not even listed in the PG docs and is also minor.
> It's not the same thing, and in most cases in Postgres prepared statements _do_not_ give you plan caching, because they are created for custom plans.
The default setting is auto which will cache the plan if the generic plan cost is similar to a custom one based on the 5 run heuristic. This is going to be most of the time on repeated simple statements that make up the bulk of application queries and why other database do this all the time globally without calling prepare. It is a large savings, not sure why you think this would not occur regularly and if you have any data to back this up I am sure everyone would like to see it, it would upset conventional thinking in other major commercial RDBMS’s with hard won gain over many years.
>You're also mixing up parsing and planning for some reason.
No I am not, you are obviously not comprehending what I said and cannot read the documentation I quoted which I had to repeat a second time here. I am not sure why you think I am mixing them up I was only trying to be gracious and include the other benefit of prepared statement, one of two thats left if it doesn't cache the plan, it avoids parsing which yes has a smaller impact, the third even less.
Also not everyone shares PG terminology, Oracle refers to what you call parsing as a soft parse (syntax check, semantic check) and parsing and planning as a hard parse (rewrite and optimizing, row source generation), you obviously have little experience outside of PG and seem to have a myopic view of what is possible in RDBMS systems and how these terms are used.
>Query parsing costs like 1/100 of planning, it's not nothing, but pretty close to it.
Again what is the point of a prepared statement if skipping parsing is meaningless and planning is not THE MAIN POINT?
>Even though you're just a rude nobody, it still may be useful for others, who may read this stupid conversation…
Further ad hominem and you call me rude, who are you to say this? How about you step off your high horse and learn something mr superior somebody. I was trying to debate in good faith and you insult me with zero substance, yeah this is a stupid conversation...
An awful lot of people have tried to use it as a JIT now and had to backpedal. I'm not sure how the one lead to the other but here we are.
Really amazed to see not one but several generic JIT frameworks though, no idea that was a thing.
This compares to clickhouse where it constantly uses the whole hardware. Obviously it's easier to do that on a columnar database but it seems that postgres is actively designed to _not_ saturate multiple cores, which may be a good assumption in the past but definitely isn't a good one now IMO.
* Latency. LLM responses are measured in order of 1000s of milliseconds, where this project targets 10s of milliseconds, that's off by almost two orders of magnitute.
* Determinism. LLMs are inherently non-deterministic. Even with temperature=0, slight variations of the input lead to major changes in output. You really don't want your DB to be non-deterministic, ever.
Better known as "seconds"...
This isn't true, and certainly not inherently so.
Changes to input leading to changes in output does not violate determinism.
From what I understand, in practice it often is true[1]:
Matrix multiplication should be “independent” along every element in the batch — neither the other elements in the batch nor how large the batch is should affect the computation results of a specific element in the batch. However, as we can observe empirically, this isn’t true.
In other words, the primary reason nearly all LLM inference endpoints are nondeterministic is that the load (and thus batch-size) nondeterministically varies! This nondeterminism is not unique to GPUs — LLM inference endpoints served from CPUs or TPUs will also have this source of nondeterminism.
[1]: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
"But why aren’t LLM inference engines deterministic? One common hypothesis is that some combination of floating-point non-associativity and concurrent execution leads to nondeterminism based on which concurrent core finishes first."
From https://thinkingmachines.ai/blog/defeating-nondeterminism-in...