> However, standard LLVM-based JIT is notoriously slow at compilation. When it takes tens to hundreds of milliseconds, it may be suitable only for very heavy, OLAP-style queries, in some cases.
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.
That's curious. I regularly get speed ups when moving processing from Python to postgres. At least when using indices properly and when the shift reduces the amount of data carried back and forth.
Oracle’s wasn’t but I haven’t used it in a very long time so that may not be longer be true.
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.
The "byte-code" coming from the query planner typically only has a handful of steps in a linear sequence. Joins, filters, and such. But the individual steps can be very costly.
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.
That looks interesting but it seems inefficient to put an LLM directly into the compilation pipeline, not to mention that it introduces nondeterministic behavior.
It has different limitations but inefficiency doesn't seem likely to be one of them. Did you read the Experimental Results section?
> 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
That's all great, but sadly impractical.
I looked at one of the first statements:
> GenDB is an LLM-powered agentic system that decomposes the
complex end-to-end query processing and optimization task into
a sequence of smaller and well-defined steps, where each step is
handled by a dedicated LLM agent.
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 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.
> It's very difficult to do low-latency queries if you cannot cache the compiled code
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.
If I recall research papers regarding Umbra it's also using AsmJit as a JIT backend, which means that theoretically the compilation times would be comparable if you only consider code emitting overhead.
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.
Interesting... AsmJit is pretty fast for compilation, but about 3x than sljit. The only way I can see how to make it fast enough, in theory (i.e. without slowing down point-lookup queries and such) would be to fuse planning with code generation - i.e. a single pass plan builder + compiler essentially. Not sure if Umbra tries to do that, and AsmJit is not the best choice for it anyway, but with sljit it could be on par with interpreter even for fastest queries I believe. Pretty hard (likely impossible) to implement though, planning is inherently a non-linear process...
Because pg_jitter uses AsmJit's Compiler, which also allocates registers. That's much more work than using hardcoded physical registers in SLJIT case. There is always a cost of such comfort.
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.
> I think it would be the best to start interpreting the query and start compilation in another thread
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.
It’s also common for JITs to sprout a tier and shed a tier over time, as the last and first tiers shift in cost/benefit. If the first tier works better you delay the other tiers. If the last tier gets faster (in run time or code optimization) you engage it sooner, or strip the middle tier entirely and hand half that budget to the last tier.
The idea with parallel compilation is interesting. Worth considering, in some cases. The only problem with it is the same as too much parallelization - you can exhaust your CPU resources much faster. But with some sort of smart scheduling it should work. I'll think about it, thanks!
Postgresql uses a process per connection model and it has no way to serialize a query plan to some form that can be shared between processes, so the time it takes to make the plan including JIT is very important.
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.
Sharing executable code between processes it not as easy as sharing data. AFAIK unless somethings changed recently PG shares nothing about plans between process and can't even share a cached plan between session/connections.
Executable code is literally just data that you mark as executable. It did the JIT code, and the idea that it can't then share it between processes is incomprehensible.
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.
No a lot of jitted code has pointers to addresses specific to that process which makes no sense in another process.
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 emphasis on compilation time there is because the JIT provider that comes with Postgres (LLVM-based) is broken in that particular area. But you're right, JITed code can be cached, if some conditions are met (it's position independent, for one). Not all JIT providers do that, but many do. Caching is on the table, but if your JIT-compilation takes microseconds, caching could be rather a burden in many cases. Still for some cases useful.
Might want to take a look at some research like this [1] that goes over the issues:
"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.
Yes if the client manually prepares the statement it will be cached for just that connection because in PG a connection is a process, but it won't survive from one connection to the next even in same process.
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.
It is not always necessary to explicitly use prepared statements, though. For example, the pgx library for Go [1] and the psycopg3 library for Python [2] will automatically manage prepared statements for you.
The last time I looked into it my impression was that disabling the JIT in PostgreSQL was the better default choice. I had a massive slowdown in some queries, and that doesn't seem to be an entirely unusual experience. It does not seem worth it to me to add such a large variability to query performance by default. The JIT seemed like something that could be useful if you benchmark the effect on your actual queries, but not as a default for everyone.
Postgres caches query plans too, the problem is you can only cache what you can share, and if your planner works well, you can share very little, there can be a lot of unique plans even for the same query
No it cannot cache query plans between processes (connections) and the only way it can cache in the same process in the same connection is by the client manually preparing it, this was how the big boys did it 30 years ago, not anymore.
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.
I didn't say it can cache between processes. The problem is not caching between processes, it's that caching itself is not very useful, because the planner creates different plans for different input parameters of the same query in the general case. So you can reliably cache plans only for the same sets of parameters. Or you can cache generic plans, which Postgres already does as well (and sharing that cache won't solve much of the problem too).
Most systems submit many of the same queries over and over again.
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.
I’m always surprised to learn LLVM is so slow given that was one of the original motivations for developing it. I don’t know if that’s down to feature creep or intrinsic complexity being higher than people presumed was the case for GCC.
It's a compiler backend for programming languages not a runtime JIT compiler. Especially inside a DBMS a lot of the assumptions it was built with don't hold. Some people in DBMS world (mostly at TUM with Umbra/CedarDB) have written their own and others tried multi pass approaches where you have an interpreter first then a more optimised LLVM pass later.
It was intended to solve the problem of interactive coding sessions such as with Language Servers, which GCC utterly fails at (because what we think of as modern IDEs did not exist in 1990).
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.
Most databases in practice are sub-terabyte and even sub-100Gb, their active dataset is almost fully cached. For most databases I worked with, cache hit rate is above 95% and for almost all of them it's above 90%. In that situation, most queries are CPU-bound. It's completely different from typical OLAP in this sense.
Definitely. If you're doing regular queries with filters on jsonb columns, having the index directly on the JSON paths is really powerful. If I have a jsonb filter in the codebase at all, it probably needs an index, unless I know the result set is already very small.
Yeah, the other problem is I've really struggled to have postgres use multiple threads/cores on one query. Often maxes out one CPU thread while dozens go unused. I constantly have to fight loads of defaults to get this to change and even then I never feel like I can get it working quite right (probably operator error to some extent).
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.
Have you tested this under high concurrency with lots of short OLTP queries? I’m curious whether the much faster compile time actually moves the point where JIT starts paying off, or if it’s still mostly useful for heavier queries.
It's not useful for sub-millisecond queries like point lookups, or other simple ones that process only a few records. sljit option starts to pay off when you process (not necessarily return) hundreds of records. The more - the better. I'm still thinking about a caching option, that will allow to lift this requirement somewhat - for cached plans. For non-cached ones it will stay.
> By default, jit_above_cost parameter is set to a very high number (100'000). This makes sense for LLVM, but doesn't make sense for faster providers. It's recommended to set this parameter value to something from ~200 to low thousands for pg_jitter (depending on what specific backend you use and your specific workloads).
Two things are holding back current LLM-style AI of being of value here:
* 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.
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.
"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."
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?