Обсуждение: Refactor query normalization into core query jumbling

Hi,

A while back, there was a discussion about moving the query
normalization code out of pg_stat_statements and into core
query jumbling [0]. This conversation also led to hardening
the hooks (currently only post_parse_analyze) that receive
JumbleState [1].

For the first point, `generate_normalized_query` takes the query
string and the jumbleState with constant locations from core
query jumbling. There is nothing uniquely special about
pg_stat_statements in this code, and it can be used by other extensions.
Extensions could even pass their own query string and JumbleState
if they choose. Patch 0001 moves this function and its helpers to
queryjumblefuncs.c.

A Github search also shows that there are quite a few extensions
that may be copying this code [2] [3], which means they will lose
out on potential improvements and fixes.

As part of this patch, the moved code itself did not change,
but I did improve a comment:

```
* If query_loc > 0, then "query" has been advanced by that much compared to
* the original string start, so we need to translate the provided locations
* to compensate.  (This lets us avoid re-scanning statements before the one
* of interest, so it's worth doing.)
```

This comment appeared at the top of generate_normalized_query and
fill_in_constant_lengths. I simplified it and moved a shortened
version inside the functions.

Also, I did not think `fill_in_constant_lengths` should be a global
function, as I cannot think of a good reason it would be used
on its own, though someone may have a different opinion there.

For the second point, since JumbleState can be shared by multiple
extensions, hooks should receive it as a const pointer. This
signals read-only intent and prevents extensions from
accidentally modifying it through the hook. This change is in
0002. This does mean that extensions will need to update their
hooks, but I do not see that as an issue for a major version.

Thoughts?

[0] https://postgr.es/m/aQA9v9nLu5qsX8IE%40paquier.xyz
[1] https://postgr.es/m/202510281023.4u5aszccvsct%40alvherre.pgsql
[2] https://github.com/search?q=fill_in_constant_lengths&type=code
[3] https://github.com/search?q=generate_normalized_query&type=code

--
Sami Imseih
Amazon Web Services (AWS)

Hi,

On Thu, Dec 18, 2025 at 06:44:17PM -0600, Sami Imseih wrote:
> For the second point, since JumbleState can be shared by multiple
> extensions, hooks should receive it as a const pointer. This
> signals read-only intent and prevents extensions from
> accidentally modifying it through the hook. This change is in
> 0002. This does mean that extensions will need to update their
> hooks, but I do not see that as an issue for a major version.

I can see that 0001 is doing:

-static void fill_in_constant_lengths(JumbleState *jstate, const char *query,

and then:

+static void
+fill_in_constant_lengths(const JumbleState *jstate, const char *query,

Should the const addition be in 0002? But...

While looking at fill_in_constant_lengths(), I can see that it is doing:

locs = jstate->clocations;

and then things like:

locs[i].length = -1

or

locs[i].length = strlen(yyextra.scanbuf + loc)

While this is technically correct so the compiler does not complain (because
clocations is a non const pointer in JumbleState and the added const does not
apply to what clocations points to), I think that adding const here is misleading.

Indeed, the function clearly modifies data accessible through the parameter.

Thoughts?

Regards,

-- 
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com

> While this is technically correct so the compiler does not complain (because
> clocations is a non const pointer in JumbleState and the added const does not
> apply to what clocations points to), I think that adding const here is misleading.

Yes, I am not happy with this. I initially thought it would be OK to
modify the JumbleState as long as it is done in a core function, but
after much thought, this is neither a good idea nor safe. If a pointer
is marked as a Constant, we should not modify it.

So, I went back to think about this, and the core problem as I see it
is that multiple hooks on the chain can modify the constant lengths. For
example, pg_stat_statements can now modify the constant lengths in
JumbleState via fill_in_constant_lengths, and the same JumbleState can
have its constant locations modified in a different way.

At this time, constant lengths are the only part of the JumbleState that
is not set by core. So, I don't think post_parse_analyze receiving
JumbleState as a constant is the right approach, nor do we need it.

I think we should allow JumbleState to define a normalization callback,
which defaults to a core normalization function rather than an
extension specific one:

```
    jstate->normalize_query = GenerateNormalizedQuery;
```

This way, any extension that wishes to return a normalized string from
the same JumbleState can invoke this callback and get consistent results.
pg_stat_statements and other extensions with a need to normalize a query
string based on the locations of a JumbleState do not need to care about the
internals of normalization, they simply invoke the callback and
receive the final
string.

So v2-0001 implements this callback and moves the normalization logic
into core. Both changes must be done in the same patch.  The comments
are also updated where they are no longer applicable or could be improved.

One additional improvement that this patch did not include is a bool in
JumbleState that tracks whether a normalized string has already been
generated. This way, repeated calls to the callback would not need to
regenerate the string; only the first call would perform the work,
while subsequent calls could simply return the previously computed
normalized string.

I do like the simplicity of this approach and it removes pg_stat_statements
from having to own the normalization code and how well different extensions
sharing the same JumbleState can play together. Not yet sure if this is the
correct direction, and I am open to other ideas.

--
Sami Imseih
Amazon Web Services (AWS)

On Fri, Dec 19, 2025 at 03:36:16PM -0600, Sami Imseih wrote:
> This way, any extension that wishes to return a normalized string from
> the same JumbleState can invoke this callback and get consistent results.
> pg_stat_statements and other extensions with a need to normalize a query
> string based on the locations of a JumbleState do not need to care about the
> internals of normalization, they simply invoke the callback and
> receive the final
> string.

Hmm.  I did not wrap completely my head with your problem, but,
assuming that what you are proposing goes in the right direction, I am
wondering if we should not expose a bit more the jumble query APIs so
as the normal default callback can be reused by out-of-core rather
than hide it entirely.  This would mean exposing
GenerateNormalizedQuery(), which also giving a way for callers of
JumbleQuery() to pass down a custom callback?  This would imply
thinking harder about the initialization state we expect in the
structure, but I think that we should try to design things so as
extensions do not need to copy-paste more code from the core tree at
the end, just less of it.

Of course, this sentence is written with the same line of thoughts as
previously mentioned in the other thread we have discussed: extensions
should not be allowed to update a JumbleState after it's been set by
the backend code, so as once the same JumbleState pointer is passed
down across multiple extensions they don't get confused.  If an
extension wants to use their own policy within the JumbleState, they
had better recreate a new independent one if they are unhappy about
has been generated previously.
--
Michael

> > This way, any extension that wishes to return a normalized string from
> > the same JumbleState can invoke this callback and get consistent results.
> > pg_stat_statements and other extensions with a need to normalize a query
> > string based on the locations of a JumbleState do not need to care about the
> > internals of normalization, they simply invoke the callback and
> > receive the final
> > string.
>
> Hmm.  I did not wrap completely my head with your problem, but,
> assuming that what you are proposing goes in the right direction,

The first goal is to move all query-normalization-related infrastructure
that pg_stat_statements (and other extensions) rely on into core, so
extensions no longer need to copy or reimplement normalization logic and
can all depend on a single, shared implementation.

In addition, query normalization necessarily modifies JumbleState (to
record constant locations and lengths). This responsibility should not
fall to extensions and should instead be delegated to core. I will argue
that the current design, in which extensions handle this directly, is a
layering violation.

As a first step, we can move generate_normalized_query to core as a global
function, allowing extensions to simply call it.

> I am wondering if we should not expose a bit more the jumble query APIs so
> as the normal default callback can be reused by out-of-core rather
> than hide it entirely.  This would mean exposing
> GenerateNormalizedQuery(), which also giving a way for callers of
> JumbleQuery() to pass down a custom callback?  This would imply
> thinking harder about the initialization state we expect in the
> structure, but I think that we should try to design things so as
> extensions do not need to copy-paste more code from the core tree at
> the end, just less of it.

... and this will be taking the next step which is providing callbacks
and making
more jumbling utilities global. This will require more discussion, but I
would think we would expose InitJumble() and it will do the bare minimum
to initialize a JumbleState, and some fields that can define callbacks after
the fact. There will be a callback for a normalization function and a
callback function that will allow the user to implement jumbling functions
for nodes that are currently not included in queryjumblefuncs.switch.c, or
perhaps they can override the existing logic in this generated file.

> Of course, this sentence is written with the same line of thoughts as
> previously mentioned in the other thread we have discussed: extensions
> should not be allowed to update a JumbleState after it's been set by
> the backend code, so as once the same JumbleState pointer is passed
> down across multiple extensions they don't get confused.  If an
> extension wants to use their own policy within the JumbleState, they
> had better recreate a new independent one if they are unhappy about
> has been generated previously.

Yes, correct. If we provide the interface to create an additional JumbleState,
they can create an independent state.

For this thread, I would like to focus on the first goal.

What do you think?

--
Sami Imseih
Amazon Web Services (AWS)

Hi,

Though this may be tangential to the current topic, I have long been wanting to revise the two instances of
`Assert(len_to_wrt>= 0);`
 
in the code to the implementation below. Would you kindly advise if this modification is worthwhile?

```
    if (len_to_wrt > 0)
    {
        memcpy(norm_query + n_quer_loc, query + quer_loc, len_to_wrt);
        n_quer_loc += len_to_wrt;
    }
```

--
Regards,
Man Zeng
www.openhalo.org

> Though this may be tangential to the current topic, I have long been wanting to revise the two instances of
`Assert(len_to_wrt>= 0);`
 
> in the code to the implementation below. Would you kindly advise if this modification is worthwhile?
>
> ```
>         if (len_to_wrt > 0)
>         {
>                 memcpy(norm_query + n_quer_loc, query + quer_loc, len_to_wrt);
>                 n_quer_loc += len_to_wrt;
>         }
> ```

I am not sure I like this idea. The len_to_wrt being less than 0
indicates a bug which will be hidden behind such a loop. I
think it's better to keep the Assert as-is.

> > > This way, any extension that wishes to return a normalized string from
> > > the same JumbleState can invoke this callback and get consistent results.
> > > pg_stat_statements and other extensions with a need to normalize a query
> > > string based on the locations of a JumbleState do not need to care about the
> > > internals of normalization, they simply invoke the callback and
> > > receive the final
> > > string.
> >
> > Hmm.  I did not wrap completely my head with your problem, but,
> > assuming that what you are proposing goes in the right direction,
>
> The first goal is to move all query-normalization-related infrastructure
> that pg_stat_statements (and other extensions) rely on into core, so
> extensions no longer need to copy or reimplement normalization logic and
> can all depend on a single, shared implementation.
>
> In addition, query normalization necessarily modifies JumbleState (to
> record constant locations and lengths). This responsibility should not
> fall to extensions and should instead be delegated to core. I will argue
> that the current design, in which extensions handle this directly, is a
> layering violation.
>
> As a first step, we can move generate_normalized_query to core as a global
> function, allowing extensions to simply call it.

v3 implements this approach without a callback. This establishes a clear
boundary: core owns JumbleState modifications, extensions consume the
results through the API.

I kept the post_parse_analyze hook signature unchanged since
GenerateNormalizedQuery still needs to modify JumbleState
(to populate constant lengths). Making it const would be misleading.

For the second part of making more jumbling utilities global, this can
be taken up in another thread. I am now thinking we would be better
off actually taking all the generic jumbling routines and separating
them into their own C file. So we can have a new file like primjumble.c
which will have all the "primitive" jumbling utilities, and queryjumblefuncs.c
can worry about its core business of jumbling a Query tree. Anyhow,
these are just some high level thoughts on this part for now.

--
Sami Imseih
Amazon Web Services (AWS)

> I am not sure I like this idea. The len_to_wrt being less than 0
> indicates a bug which will be hidden behind such a loop. I
> think it's better to keep the Assert as-is.

Well, my main concern is that assertions are unlikely to be enabled in production environments. 
Also, the new patch looks really good. Besides, it seems you accidentally forgot to cc Michael.

--
Regards,
Man Zeng
www.openhalo.org

On Wed, Dec 24, 2025 at 10:50:16AM +0800, zengman wrote:
> Besides, it seems you accidentally forgot to cc Michael.

This is not a problem.  I still get the messages by being registered
to this list :D
--
Michael

On Tue, Dec 23, 2025 at 10:35:18AM -0600, Sami Imseih wrote:
> > Though this may be tangential to the current topic, I have long been wanting to revise the two instances of
`Assert(len_to_wrt>= 0);` 
> > in the code to the implementation below. Would you kindly advise if this modification is worthwhile?
> >
> > ```
> >         if (len_to_wrt > 0)
> >         {
> >                 memcpy(norm_query + n_quer_loc, query + quer_loc, len_to_wrt);
> >                 n_quer_loc += len_to_wrt;
> >         }
> > ```
>
> I am not sure I like this idea. The len_to_wrt being less than 0
> indicates a bug which will be hidden behind such a loop. I
> think it's better to keep the Assert as-is.

This set of asserts are critical to keep.  An incorrect computation is
a synonym of an out-of-bound write in this case, which would classify
any problem as something that could be worth a CVE.

> v3 implements this approach without a callback. This establishes a clear
> boundary: core owns JumbleState modifications, extensions consume the
> results through the API.
>
> I kept the post_parse_analyze hook signature unchanged since
> GenerateNormalizedQuery still needs to modify JumbleState
> (to populate constant lengths). Making it const would be misleading.
>
> For the second part of making more jumbling utilities global, this can
> be taken up in another thread. I am now thinking we would be better
> off actually taking all the generic jumbling routines and separating
> them into their own C file. So we can have a new file like primjumble.c
> which will have all the "primitive" jumbling utilities, and queryjumblefuncs.c
> can worry about its core business of jumbling a Query tree. Anyhow,
> these are just some high level thoughts on this part for now.

Hmm.  Moving from pgss to the core backup the code that updates the
clocations of an existing JumbleState does not solve the issue that
we should not modify the internals of the JumbleState computed.  So
this does not move the ball at all.

The updates of clocations depend on GenerateNormalizedQuery(), which
depends on the pre-work done in CleanQuerytext() to get a query string
and its location for a multi-query string case.  If we really want to
make a JumbleState not touchable at all, we could either push more
work of CleanQuerytext() and GenerateNormalizedQuery() into core.

Or  a second thing that we can do, which would be actually much
simpler, is to rework entirely fill_in_constant_lengths() so as we
generate a *copy* of the location data PGSS wants rather than force it
into a JumbleState that would be pushed across more stacks of the
post-parse-analyze hook.  Doing that would allow us to pull the plug
into making JumbleState and the original copies of clocations a set of
consts when given to extensions.
--
Michael

> > For the second part of making more jumbling utilities global, this can
> > be taken up in another thread. I am now thinking we would be better
> > off actually taking all the generic jumbling routines and separating
> > them into their own C file. So we can have a new file like primjumble.c
> > which will have all the "primitive" jumbling utilities, and queryjumblefuncs.c
> > can worry about its core business of jumbling a Query tree. Anyhow,
> > these are just some high level thoughts on this part for now.
>
> Hmm.  Moving from pgss to the core backup the code that updates the
> clocations of an existing JumbleState does not solve the issue that
> we should not modify the internals of the JumbleState computed.  So
> this does not move the ball at all.

This is for the point "I am wondering if we should not expose a bit more
the jumble query APIs" [0], which as I mention is a separate topic and is
not about extensions being able to modify JumbleState.

> The updates of clocations depend on GenerateNormalizedQuery(), which
> depends on the pre-work done in CleanQuerytext() to get a query string
> and its location for a multi-query string case.  If we really want to
> make a JumbleState not touchable at all, we could either push more
> work of CleanQuerytext() and GenerateNormalizedQuery() into core.

When you say “push more work into core,” I understand that to mean this work
should occur during jumbling. If so, there are several complications.

1/ CleanQueryText() requires access to the query text, which we do not have
during core jumbling as of 2ecbb0a4935.

2/ GenerateNormalizedQuery() should be invoked on demand by the extension
that needs the normalized string, for example when pg_stat_statements needs
to store a query string.

Both operations are expensive, especially GenerateNormalizedQuery(). Doing
this work on every JumbleQuery() would introduce unnecessary overhead.

> Or  a second thing that we can do, which would be actually much
> simpler, is to rework entirely fill_in_constant_lengths() so as we
> generate a *copy* of the location data PGSS wants rather than force it
> into a JumbleState

yes, that's an idea that did cross my mind. I have hesitation of copying
clocations, but that may not be such a big deal, since it will only be
occurring on demand when the extension requests a normalized string.

> that would be pushed across more stacks of the
> post-parse-analyze hook.  Doing that would allow us to pull the plug
> into making JumbleState and the original copies of clocations a set of
> consts when given to extensions.

Yes correct. The hook signature will turn into, so all extensions now
just have a constant pointer to the jumblestate. They can of course
work hard to cast away constants, but at that point, they are doing
things the wrong way.

```
 typedef void (*post_parse_analyze_hook_type) (ParseState *pstate,

        Query *query,

        const JumbleState *jstate);
```

This of course will be a big breaking change to all the extensions out
there using this hook. This is not just about a signature change of
the hook, but an author now has to figure out how to deal with a
constant JumbleState in their code, which may not be trivial.

My proposal in v3 was a bit more softer, and keeps the hook
backwards compatible. Basically, we keep JumbleState a non-constant,
but provide core APIs for any  operation, such as generate_normalized_query,
that needs to modify the state. So, my approach was not about enforcing a
read-only JumbleState, but about providing the API to dissuade an author
from modifying a JumbleState.
If we need a stronger API contract, then we can go with the hook receving
JumbleState as a constant and implement the copy of clocations to return
back to extensions that need to normalize a query. We also have to keep
in mind that if there are future requirements where the state has to be
modified by an extension, we will need to implement more copy functions
for those field.

[0] https://www.postgresql.org/message-id/aUXeTEpSymo6n7lF%40paquier.xyz

--
Sami Imseih
Amazon Web Services (AWS)

On Mon, Dec 29, 2025 at 06:34:18PM -0800, Lukas Fittl wrote:
> I noticed pg_hint_plan checks the JumbleState argument to determine whether
> or not to run an early check of the hint logic. I'm a little
> confused by this (and if this is about query IDs being available,
> couldn't it just check the  Query having a non-zero queryID?), but
> FWIW: [3]

See also the code before 32ced2e13e75.  We wanted a JumbleState in
this case to be able to generate a normalized query for the hint
table.  Since this commit we only rely on the query ID in the hint
table, so it would not matter for pg_hint_plan if the JumbleState is
removed from this portions of the hooks.

Saying that, yes I think that you are right, we should be able to
remove this dependency on JumbleState in
pg_hint_plan_post_parse_analyze() and rely on the query ID.  I'm
writing down a note about cleaning that on the HEAD branch of the
module..
--
Michael

Hi,

On Tue, Dec 23, 2025 at 10:35:18AM -0600, Sami Imseih wrote:
> v3 implements this approach without a callback. This establishes a clear
> boundary: core owns JumbleState modifications, extensions consume the
> results through the API.
> 

Thanks for the new patch version.

Some random comments:

=== 1

+       SetConstantLengths((JumbleState *) jstate, query, query_loc);

This cast seems unnecessary.

=== 2

+CompLocation(const void *a, const void *b)

In the commit message I can see "Functions are renamed to match core naming
conventions" but wasn't comp_location() better? 

=== 3

+                       /*
+                        * generate the normalized query. Note that the normalized
+                        * representation may well vary depending on just which
+                        * "equivalent" query is used to create the hashtable entry. We
+                        * assume this is OK.
+                        */
+                       norm_query = GenerateNormalizedQuery(jstate, query,

Should part of this comment be on top of the GenerateNormalizedQuery()
definition instead?

Regards,

-- 
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com

> === 1
>
> +       SetConstantLengths((JumbleState *) jstate, query, query_loc);
>
> This cast seems unnecessary.

Right. This is unnecessary. I will remove in the next iteration.

> === 2
>
> +CompLocation(const void *a, const void *b)
>
> In the commit message I can see "Functions are renamed to match core naming
> conventions" but wasn't comp_location() better?

Not sure if it's better or worse. I aimed for consistency here and
used pascal case to
match.


> === 3
>
> +                       /*
> +                        * generate the normalized query. Note that the normalized
> +                        * representation may well vary depending on just which
> +                        * "equivalent" query is used to create the hashtable entry. We
> +                        * assume this is OK.
> +                        */
> +                       norm_query = GenerateNormalizedQuery(jstate, query,
>
> Should part of this comment be on top of the GenerateNormalizedQuery()
> definition instead?

No, because this talks about the hashtable entry which is specific to
pg_stat_statements.

--
Sami Imseih
Amazon Web Services (AWS)

>> backwards compatible. Basically, we keep JumbleState a non-constant,
>> but provide core APIs for any operation, such as
>> generate_normalized_query,
>> that needs to modify the state. So, my approach was not about enforcing a
>> read-only JumbleState, but about providing the API to dissuade an author
>> from modifying a JumbleState.

> Given the lack of public APIs to modify JumbleState today, I don't see how
> an extension would do
> modifications in a meaningful way, short of copying the code. I think we
> should be a bit bolder here in
> enforcing a convention, either clearly making it read-only or dropping the
> argument again.

Based on the discussion so far I am leaning towards making JumbleState
read-only as described here [0]. I don't see how we can drop JumbleState
completely from hooks, since normalization needs to occur on-demand
by the extension.

> 1) An extension that wants to display normalized query strings
>
> This seems to be the biggest kind of what I can find with code search.
> Extensions like pg_stat_monitor [1], that
> want to do something like pg_stat_statements, and have to copy a bunch of
> normalization code today that is 1:1 what
> Postgres does. Such extensions don't need the JumbleState argument if they
> can get the normalized text directly.

Yes, I don't know how that's possible; besides generating the normalized
string during JumbleQuery and making it available to post_parse_analyze
hook ( and other executor hooks ). But this also means we are incurring
the normalization overhead for every execution.

> 2) An extension that wants to capture parameter values
>
> Some extensions may want to remember additional context for normalized
> queries, like pg_tracing's logic for
> optionally passing parameter values (post normalization) in the trace
> context [2]. If we kept passing a read-only
> JumbleState then such extensions could presumably still get this, but I
> wonder if it wouldn't be better for core to
> have a helper for this?

This could be like a core GenerateNormalizedQuery which can optionally
track the constant values. That will be an enhancement to normalization
and a new requirement.

[0] https://www.postgresql.org/message-id/CAA5RZ0sbWmqdUBFo8JXMJe72pnwjxVY58htJ6pKbwnyQuRctQw%40mail.gmail.com

--
Sami Imseih
Amazon Web Services (AWS)

> >> backwards compatible. Basically, we keep JumbleState a non-constant,
> >> but provide core APIs for any operation, such as
> >> generate_normalized_query,
> >> that needs to modify the state. So, my approach was not about enforcing a
> >> read-only JumbleState, but about providing the API to dissuade an author
> >> from modifying a JumbleState.
>
> > Given the lack of public APIs to modify JumbleState today, I don't see how
> > an extension would do
> > modifications in a meaningful way, short of copying the code. I think we
> > should be a bit bolder here in
> > enforcing a convention, either clearly making it read-only or dropping the
> > argument again.
>
> Based on the discussion so far I am leaning towards making JumbleState
> read-only as described here [0]. I don't see how we can drop JumbleState
> completely from hooks, since normalization needs to occur on-demand
> by the extension.

Here is v4.

0001 - addresses the comments made by Bertrand.
0002 - makes JumbleState a constant when passed to post_parse_analyze
and updates all downstream code that consume the JumbleState. This
means we now need to copy the locations from JState into a local/temporary
array when generating the normalized string.

--
Sami Imseih
Amazon Web Services (AWS)

On Tue, Dec 30, 2025 at 01:31:30PM +0900, Michael Paquier wrote:
> Saying that, yes I think that you are right, we should be able to
> remove this dependency on JumbleState in
> pg_hint_plan_post_parse_analyze() and rely on the query ID.  I'm
> writing down a note about cleaning that on the HEAD branch of the
> module..

For the note, I have looked at that again and this improvement is
right, so I have pushed something to simplify the code:
https://github.com/ossc-db/pg_hint_plan/commit/3a63a56740e48e1d205294d5df6e198716718882

Thanks for the suggestion, Lukas.
--
Michael