PROPERTY GRAPH pg_dump ACL minimization

Noah Misch <noah@leadboat.com> writes:

Most of the patch bulk (modest as it is) exists to keep support for dumping
from beta1. I'm not sure whether it was worth bothering. Breaking dump from
a beta is without precedent known to me, so I just erred on the side of not
breaking it. If we were to decide pg_dump could drop support for betas, I'd
be fine with that.

This entails a catversion bump on the v19 branch.

Those points are not unrelated. If you bump catversion then beta
testers must use pg_upgrade to get from beta1 to beta2, so you should
not drop support for dumping from beta1.

I could agree with dropping that support after beta2, though. That'd
imply having to update via beta2 to beta3 or later, but I doubt those
hardy enough to test beta1 would have a problem with that.

If the master branch
already has a post-branch catversion bump by then, the two catversions should
remain distinct. I'll use yyyymmdd1 for v19 and yyyymmdd2 for master.

Check.

(I didn't read the patch, just responded to your commentary.)

regards, tom lane

Thanks Noah for the patch and the report.

On Tue, Jun 30, 2026 at 8:41 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:

Noah Misch <noah@leadboat.com> writes:

Most of the patch bulk (modest as it is) exists to keep support for dumping
from beta1. I'm not sure whether it was worth bothering. Breaking dump from
a beta is without precedent known to me, so I just erred on the side of not
breaking it. If we were to decide pg_dump could drop support for betas, I'd
be fine with that.

This entails a catversion bump on the v19 branch.

Those points are not unrelated. If you bump catversion then beta
testers must use pg_upgrade to get from beta1 to beta2, so you should
not drop support for dumping from beta1.

I could agree with dropping that support after beta2, though. That'd
imply having to update via beta2 to beta3 or later, but I doubt those
hardy enough to test beta1 would have a problem with that.

Should there be two commits one which will be reverted post beta2 and
one which will stay after beta2?

If the master branch
already has a post-branch catversion bump by then, the two catversions should
remain distinct. I'll use yyyymmdd1 for v19 and yyyymmdd2 for master.

Check.

(I didn't read the patch, just responded to your commentary.)

I reviewed the patch.

- tblinfo[i].dacl.acldefault = pg_strdup(PQgetvalue(res, i, i_acldefault));
+ /* acldefault computed below */

Rather than spacially separating the acldefault computation, can we
just write a function to compute the acldefault for a given relkind
and owner, and call that function here?

Did you consider writing a test for the same?

Other than that the patch looks good.

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

1. ExecGrant_Relation: I think we should clip the extra privileges
with a warning when GRANT ... TABLE syntax is used to grant privileges
on a property graph, just like sequences. To me it looks like we
should prohibit GRANT ... TABLE on property graph altogether. But
haven't done so to keep it in sync with sequences. The backward
compatibility comment, "For backward compatibility, just ... " should
not be applicable in case of property graph since we can introduce
whatever behaviour we expect from GRANT ... TABLE right from the first
release which introduced property graph. But I am not sure if that's
the only backward compatibility we are talking about here. Those
commits go more than a few decades back and commit message itself
doesn't help me much. Maybe someone with a better historical
perspective may help. I have also added a test scenario for a
non-property graph privilege to be added using GRANT ... TABLE syntax.

The second change in this function seems necessary but without it, I
couldn't find a visible bug. Mostly it's masked because the privileges
available on a table are a superset of privileges available on a
property graph.

Now that the function handles property graph separately, its prologue
needs to change. I think we should mention property graph along with
sequences as done in the patch OR just mention "all types of objects
in pg_class."

2. pg_class_aclmask_ext(): this seems to be another omission. Probably
innocuous since we will test only SELECT privileges on a property
graph and ignore other default table privileges.

3. pg_default_acl: doesn't need any update since property graph is not
an object listed in the types of objects for which the user is allowed
to specify default permissions through pg_default_acl.

All other places which handle OBJECT_SEQUENCE also handle OBJECT_PROPGRAPH.

I also notice that information_schema.pg_propgraph_privileges shows
only privileges of type "SELECT" so we wouldn't be able to notice a
privilege type other than SELECT being granted on a property graph
through information_schema. But a similar filtering exists in the view
information_schema.table_privileges. So it looks intentional and I
didn't touch it.

--
Best Wishes,
Ashutosh Bapat

On Wed, Jul 1, 2026 at 9:51 PM Ashutosh Bapat
<ashutosh.bapat.oss@gmail.com> wrote:

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

1. ExecGrant_Relation: I think we should clip the extra privileges
with a warning when GRANT ... TABLE syntax is used to grant privileges
on a property graph, just like sequences. To me it looks like we
should prohibit GRANT ... TABLE on property graph altogether. But
haven't done so to keep it in sync with sequences. The backward
compatibility comment, "For backward compatibility, just ... " should
not be applicable in case of property graph since we can introduce
whatever behaviour we expect from GRANT ... TABLE right from the first
release which introduced property graph. But I am not sure if that's
the only backward compatibility we are talking about here. Those
commits go more than a few decades back and commit message itself
doesn't help me much. Maybe someone with a better historical
perspective may help. I have also added a test scenario for a
non-property graph privilege to be added using GRANT ... TABLE syntax.

Since property graphs share the namespace with regular tables, I think
GRANT ... TABLE should be supported on property graphs, but restrict
it to only the privileges applicable to property graphs. Done that way
in the attached patch.

The second change in this function seems necessary but without it, I
couldn't find a visible bug. Mostly it's masked because the privileges
available on a table are a superset of privileges available on a
property graph.

This change is needed so that we can provide a correct error message.

Here's a revised patch set.
0010 is your patch without any changes
0011 is my changes described above.

--
Best Wishes,
Ashutosh Bapat

On Wed, Jul 01, 2026 at 09:51:13PM +0530, Ashutosh Bapat wrote:

On Tue, Jun 30, 2026 at 8:41 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:

Noah Misch <noah@leadboat.com> writes:

Most of the patch bulk (modest as it is) exists to keep support for dumping
from beta1. I'm not sure whether it was worth bothering. Breaking dump from
a beta is without precedent known to me, so I just erred on the side of not
breaking it. If we were to decide pg_dump could drop support for betas, I'd
be fine with that.

This entails a catversion bump on the v19 branch.

Those points are not unrelated. If you bump catversion then beta
testers must use pg_upgrade to get from beta1 to beta2, so you should
not drop support for dumping from beta1.

I could agree with dropping that support after beta2, though. That'd
imply having to update via beta2 to beta3 or later, but I doubt those
hardy enough to test beta1 would have a problem with that.

I agree dropping 19beta1 dump support after beta2 would be reasonable, and I
hadn't thought of doing so. However, ...

Should there be two commits one which will be reverted post beta2 and
one which will stay after beta2?

... since the sole benefit would be removing ~25 lines, I don't think a
followup patch is worth it. I wouldn't object if someone wants to remove that
code later.

If the master branch
already has a post-branch catversion bump by then, the two catversions should
remain distinct. I'll use yyyymmdd1 for v19 and yyyymmdd2 for master.

Check.

(I didn't read the patch, just responded to your commentary.)

I reviewed the patch.

- tblinfo[i].dacl.acldefault = pg_strdup(PQgetvalue(res, i, i_acldefault));
+ /* acldefault computed below */

Rather than spacially separating the acldefault computation, can we
just write a function to compute the acldefault for a given relkind
and owner, and call that function here?

Such a function would have just one caller, and that would even further
spatially separate the code lines defining the computation. Either way is
fine, but I'm not inclined to change to that.

Did you consider writing a test for the same?

Somewhat. Long-term, I think a more general test would have a place.

Other than that the patch looks good.

Thanks for reviewing.

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

This probably calls for its own thread; feel free to fork the thread for any
followup on that.

On Fri, Jul 03, 2026 at 04:09:20PM +0530, Ashutosh Bapat wrote:

On Wed, Jul 1, 2026 at 9:51 PM Ashutosh Bapat
<ashutosh.bapat.oss@gmail.com> wrote:

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

1. ExecGrant_Relation: I think we should clip the extra privileges
with a warning when GRANT ... TABLE syntax is used to grant privileges
on a property graph, just like sequences. To me it looks like we
should prohibit GRANT ... TABLE on property graph altogether. But
haven't done so to keep it in sync with sequences. The backward
compatibility comment, "For backward compatibility, just ... " should
not be applicable in case of property graph since we can introduce
whatever behaviour we expect from GRANT ... TABLE right from the first
release which introduced property graph. But I am not sure if that's
the only backward compatibility we are talking about here. Those
commits go more than a few decades back and commit message itself
doesn't help me much. Maybe someone with a better historical
perspective may help. I have also added a test scenario for a
non-property graph privilege to be added using GRANT ... TABLE syntax.

Since property graphs share the namespace with regular tables, I think
GRANT ... TABLE should be supported on property graphs, but restrict
it to only the privileges applicable to property graphs.

I don't have a strong opinion on that, but I likely would have chosen to block
GRANT TABLE on a propgraph. The backward compatibility argument written for
SEQUENCE likely means that someone noticed GRANT TABLE worked on sequences and
decided both that it was a mistake and that reversing the mistake would be a
cure worse than the disease. There was a rebuttable presumption that when we
add a new grantable object with its own GRANT subtype, older GRANT subtypes
should reject that object.

2. pg_class_aclmask_ext(): this seems to be another omission. Probably
innocuous since we will test only SELECT privileges on a property
graph and ignore other default table privileges.

Makes sense.

Noah Misch <noah@leadboat.com> writes:

On Wed, Jul 01, 2026 at 09:51:13PM +0530, Ashutosh Bapat wrote:

Since property graphs share the namespace with regular tables, I think
GRANT ... TABLE should be supported on property graphs, but restrict
it to only the privileges applicable to property graphs.

I don't have a strong opinion on that, but I likely would have chosen to block
GRANT TABLE on a propgraph. The backward compatibility argument written for
SEQUENCE likely means that someone noticed GRANT TABLE worked on sequences and
decided both that it was a mistake and that reversing the mistake would be a
cure worse than the disease.

My recollection is that there was an intentional policy change.
Originally the idea was "why make people be careful about which
kind of relation they're granting on?". The arguments made
against that included:

* It's exposing an implementation detail, namely that sequences
and tables live in the same catalog. Admittedly that detail is
also exposed by the fact that they can't share a name.

* It doesn't comport very well with the fact that the sets of
possible privileges are different.

* It doesn't obey the SQL standard (I think, maybe someone will
correct me).

But you are entirely right that we felt that disallowing what used
to work was worse than leaving it alone. We need not duplicate that
mistake for a new kind of relation, and should not.

regards, tom lane

On Sat, Jul 4, 2026 at 8:16 PM Noah Misch <noah@leadboat.com> wrote:

On Wed, Jul 01, 2026 at 09:51:13PM +0530, Ashutosh Bapat wrote:

On Tue, Jun 30, 2026 at 8:41 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:

Noah Misch <noah@leadboat.com> writes:

Most of the patch bulk (modest as it is) exists to keep support for dumping
from beta1. I'm not sure whether it was worth bothering. Breaking dump from
a beta is without precedent known to me, so I just erred on the side of not
breaking it. If we were to decide pg_dump could drop support for betas, I'd
be fine with that.

This entails a catversion bump on the v19 branch.

Those points are not unrelated. If you bump catversion then beta
testers must use pg_upgrade to get from beta1 to beta2, so you should
not drop support for dumping from beta1.

I could agree with dropping that support after beta2, though. That'd
imply having to update via beta2 to beta3 or later, but I doubt those
hardy enough to test beta1 would have a problem with that.

I agree dropping 19beta1 dump support after beta2 would be reasonable, and I
hadn't thought of doing so. However, ...

Should there be two commits one which will be reverted post beta2 and
one which will stay after beta2?

... since the sole benefit would be removing ~25 lines, I don't think a
followup patch is worth it. I wouldn't object if someone wants to remove that
code later.

As long as the code that needs to be removed after beta2 is clear, I
am fine. If somebody other than the ones involved here want to remove
the code, we should make it easy for them to do so.

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

This probably calls for its own thread; feel free to fork the thread for any
followup on that.

Will do that soon.

--
Best Wishes,
Ashutosh Bapat

On Sat, Jul 4, 2026 at 9:21 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

Noah Misch <noah@leadboat.com> writes:

On Wed, Jul 01, 2026 at 09:51:13PM +0530, Ashutosh Bapat wrote:

Since property graphs share the namespace with regular tables, I think
GRANT ... TABLE should be supported on property graphs, but restrict
it to only the privileges applicable to property graphs.

I don't have a strong opinion on that, but I likely would have chosen to block
GRANT TABLE on a propgraph. The backward compatibility argument written for
SEQUENCE likely means that someone noticed GRANT TABLE worked on sequences and
decided both that it was a mistake and that reversing the mistake would be a
cure worse than the disease.

My recollection is that there was an intentional policy change.
Originally the idea was "why make people be careful about which
kind of relation they're granting on?". The arguments made
against that included:

* It's exposing an implementation detail, namely that sequences
and tables live in the same catalog. Admittedly that detail is
also exposed by the fact that they can't share a name.

* It doesn't comport very well with the fact that the sets of
possible privileges are different.

* It doesn't obey the SQL standard (I think, maybe someone will
correct me).

But you are entirely right that we felt that disallowing what used
to work was worse than leaving it alone. We need not duplicate that
mistake for a new kind of relation, and should not.

Thanks for the clarification. I will change the patch to prohibit
using GRANT ... TABLE on a property graph.

--
Best Wishes,
Ashutosh Bapat

On Mon, Jun 29, 2026 at 10:33 PM Noah Misch <noah@leadboat.com> wrote:

pg_dump doesn't do its normal ACL minimization for the new PROPERTY GRAPH
feature. Patch attached. See log message for details.

Most of the patch bulk (modest as it is) exists to keep support for dumping
from beta1. I'm not sure whether it was worth bothering. Breaking dump from
a beta is without precedent known to me, so I just erred on the side of not
breaking it. If we were to decide pg_dump could drop support for betas, I'd
be fine with that.

Thanks for catching this, Noah.

While reviewing this patch, I wondered whether quoteAclUserName()
exactly matches what the server would do. It has this dire warning:

/* This test had better match what putid() does */

It doesn't any more, quite, because quoteAclUserName() uses this test:

!isalnum((unsigned char) *src) && *src != '_'

And putid uses this test:

!is_safe_acl_char(*src, false)

The difference between the two is that is_safe_acl_char(c) will return
false for any bytes where IS_HIGHBIT_SET(c) returns true. So consider:

CREATE ROLE álvaro;
CREATE PROPERTY GRAPH herrera;
ALTER PROPERTY GRAPH herrera OWNER TO álvaro;
GRANT SELECT ON PROPERTY GRAPH herrera TO PUBLIC;

In a UTF8 database, everything is fine. But with encoding = LATIN1 and
lc_ctype = en_US.ISO8859-1, pg_class.relacl is display with quotes
around álvaro, and the string pg_dump synthesizes lacks them. This
doesn't seem to cause a functional problem, because the ACLs are not
directly compared -- they get parsed first, and that undoes the
quoting. It seems a tad fragile, maybe, but perhaps not worth worrying
about. It's also not really the fault of this patch anyway, but it was
the only thing I found while looking through this, so I figured I
would mention it.

--
Robert Haas
EDB: http://www.enterprisedb.com

On Mon, Jul 06, 2026 at 12:40:05PM -0400, Robert Haas wrote:

On Mon, Jun 29, 2026 at 10:33 PM Noah Misch <noah@leadboat.com> wrote:

pg_dump doesn't do its normal ACL minimization for the new PROPERTY GRAPH
feature. Patch attached. See log message for details.

While reviewing this patch, I wondered whether quoteAclUserName()
exactly matches what the server would do. It has this dire warning:

/* This test had better match what putid() does */

It doesn't any more, quite, because quoteAclUserName() uses this test:

!isalnum((unsigned char) *src) && *src != '_'

And putid uses this test:

!is_safe_acl_char(*src, false)

The difference between the two is that is_safe_acl_char(c) will return
false for any bytes where IS_HIGHBIT_SET(c) returns true. So consider:

CREATE ROLE álvaro;
CREATE PROPERTY GRAPH herrera;
ALTER PROPERTY GRAPH herrera OWNER TO álvaro;
GRANT SELECT ON PROPERTY GRAPH herrera TO PUBLIC;

In a UTF8 database, everything is fine. But with encoding = LATIN1 and
lc_ctype = en_US.ISO8859-1, pg_class.relacl is display with quotes
around álvaro, and the string pg_dump synthesizes lacks them. This
doesn't seem to cause a functional problem, because the ACLs are not
directly compared -- they get parsed first, and that undoes the
quoting. It seems a tad fragile, maybe, but perhaps not worth worrying
about. It's also not really the fault of this patch anyway, but it was
the only thing I found while looking through this, so I figured I
would mention it.

Thanks for reviewing and for identifying that. I think this boils down to the
comment being too dire given current use cases. It's important for both
quoteAclUserName() and putid() to quote metacharacters, but it's okay if one
of them quotes non-metacharacters that the other doesn't quote.

I guess the behavior mismatch you've identified also makes this comment wrong:

/*
* Test whether an identifier char can be left unquoted in ACLs.
*
* Formerly, we used isalnum() even on non-ASCII characters, resulting in
* unportable behavior. To ensure dump compatibility with old versions,
* we now treat high-bit-set characters as always requiring quoting during
* putid(), but getid() will always accept them without quotes.
*/
static inline bool
is_safe_acl_char(unsigned char c, bool is_getid)

It's not just "compatibility with old versions" as long as pg_dump calls to
quoteAclUserName() are still churning out affected values.

Best if we make quoteAclUserName() match putid(), so the comments can just be
correct and nobody needs to wonder about the mismatch.

Starting a new thread from [1] as suggested.

On Fri, Jul 03, 2026 at 04:09:20PM +0530, Ashutosh Bapat wrote:

On Wed, Jul 1, 2026 at 9:51 PM Ashutosh Bapat
<ashutosh.bapat.oss@gmail.com> wrote:

I wondered whether we are missing special handling for PROPGRAPH at
other places. I looked at other places where we handle OBJECT_SEQUENCE
separately in acl related files. I discovered following missing cases

1. ExecGrant_Relation: I think we should clip the extra privileges
with a warning when GRANT ... TABLE syntax is used to grant privileges
on a property graph, just like sequences. To me it looks like we
should prohibit GRANT ... TABLE on property graph altogether. But
haven't done so to keep it in sync with sequences. The backward
compatibility comment, "For backward compatibility, just ... " should
not be applicable in case of property graph since we can introduce
whatever behaviour we expect from GRANT ... TABLE right from the first
release which introduced property graph. But I am not sure if that's
the only backward compatibility we are talking about here. Those
commits go more than a few decades back and commit message itself
doesn't help me much. Maybe someone with a better historical
perspective may help. I have also added a test scenario for a
non-property graph privilege to be added using GRANT ... TABLE syntax.

Since property graphs share the namespace with regular tables, I think
GRANT ... TABLE should be supported on property graphs, but restrict
it to only the privileges applicable to property graphs.

I don't have a strong opinion on that, but I likely would have chosen to block
GRANT TABLE on a propgraph. The backward compatibility argument written for
SEQUENCE likely means that someone noticed GRANT TABLE worked on sequences and
decided both that it was a mistake and that reversing the mistake would be a
cure worse than the disease. There was a rebuttable presumption that when we
add a new grantable object with its own GRANT subtype, older GRANT subtypes
should reject that object.

2. pg_class_aclmask_ext(): this seems to be another omission. Probably
innocuous since we will test only SELECT privileges on a property
graph and ignore other default table privileges.

Makes sense.

Modified the patch to disallow GRANT ... TABLE on property graph.

The patch uses errmsg("\"%s\" is a property graph",
NameStr(pg_class_tuple->relname)) inline with other prohibited cases.
Additionally it gives errhint("Use GRANT ... ON PROPERTY GRAPH
instead.")), which may be obvious from the error message, but I
thought clarity is better than brevity.

There are some cases where we report ""%s" is not a sequence". Inline
with that we could use errmsg("\"%s\" is not a table",
NameStr(pg_class_tuple->relname)) with errhint("Use GRANT ... ON
PROPERTY GRAPH instead.")).

I prefer the first one, but I am ok with the second option as well.

--
Best Wishes,
Ashutosh Bapat