Thread: PATCH: Attempt to make dbsize a bit more consistent
Hi all, this minor patch is attempting to force the use of the tableam api in dbsize where ever it is required. Apparently something similar was introduced for toast relations only. Intuitively it seems that the distinction between atable and a toast table is not needed. This patch treats tables, toast tables and materialized views equally. Regards, //Georgios
Attachment
On Thu, Aug 27, 2020 at 9:39 AM <gkokolatos@pm.me> wrote: > > Hi all, > > this minor patch is attempting to force the use of the tableam api in dbsize where ever it is required. > > Apparently something similar was introduced for toast relations only. Intuitively it seems that the distinction betweena table and a toast table is not needed. I suspect the reason is found in the comment for table_block_relation_size(): * If a table AM uses the various relation forks as the sole place where data * is stored, and if it uses them in the expected manner (e.g. the actual data * is in the main fork rather than some other), it can use this implementation * of the relation_size callback rather than implementing its own. -- John Naylor https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
I found the function "table_relation_size" is only used by buffer
manager for "RELKIND_RELATION", "RELKIND_TOASTVALUE" and
"RELKIND_MATVIEW", i.e.
case RELKIND_RELATION:
case RELKIND_TOASTVALUE:
case RELKIND_MATVIEW:
{
/*
* Not every table AM uses BLCKSZ wide fixed size blocks.
* Therefore tableam returns the size in bytes - but
for the
* purpose of this routine, we want the number of blocks.
* Therefore divide, rounding up.
*/
uint64 szbytes;
szbytes = table_relation_size(relation, forkNum);
return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
}
So using "calculate_relation_size" and "calculate_toast_table_size" in
"calculate_table_size" is easy to understand and the original logic is
simple.
On 2020-08-27 6:38 a.m., gkokolatos@pm.me wrote:
> Hi all,
>
> this minor patch is attempting to force the use of the tableam api in dbsize where ever it is required.
>
> Apparently something similar was introduced for toast relations only. Intuitively it seems that the distinction
betweena table and a toast table is not needed. This patch treats tables, toast tables and materialized views equally.
>
> Regards,
> //Georgios
Best regards,
--
David
Software Engineer
Highgo Software Inc. (Canada)
www.highgo.ca
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Tuesday, 8 September 2020 16:49, John Naylor <john.naylor@2ndquadrant.com> wrote:
> On Thu, Aug 27, 2020 at 9:39 AM gkokolatos@pm.me wrote:
>
> > Hi all,
> > this minor patch is attempting to force the use of the tableam api in dbsize where ever it is required.
> > Apparently something similar was introduced for toast relations only. Intuitively it seems that the distinction
betweena table and a toast table is not needed.
>
> I suspect the reason is found in the comment for table_block_relation_size():
>
> - If a table AM uses the various relation forks as the sole place where data
> - is stored, and if it uses them in the expected manner (e.g. the actual data
> - is in the main fork rather than some other), it can use this implementation
> - of the relation_size callback rather than implementing its own.
Thank you for your answer and interest at the patch.
I agree with the comment above. However I do not see why it is relevant here. When issuing:
SELECT pg_table_size('foo'::regclass);
I should not have to care about the on disk layout of the relation 'foo'.
Without this patch, one will get a correct result only when 'foo' is a heap table.
For custom layouts the result can potentially be wrong.
>
> --
> John Naylor https://www.2ndQuadrant.com/
> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
>
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Tuesday, 8 September 2020 22:26, David Zhang <david.zhang@highgo.ca> wrote:
>
>
> I found the function "table_relation_size" is only used by buffer
> manager for "RELKIND_RELATION", "RELKIND_TOASTVALUE" and
> "RELKIND_MATVIEW", i.e.
>
> case RELKIND_RELATION:
> case RELKIND_TOASTVALUE:
> case RELKIND_MATVIEW:
> {
> /*
> * Not every table AM uses BLCKSZ wide fixed size blocks.
> * Therefore tableam returns the size in bytes - but
> for the
> * purpose of this routine, we want the number of blocks.
> * Therefore divide, rounding up.
> */
> uint64 szbytes;
>
> szbytes = table_relation_size(relation, forkNum);
>
> return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
> }
>
> So using "calculate_relation_size" and "calculate_toast_table_size" in
> "calculate_table_size" is easy to understand and the original logic is
> simple.
>
You are correct. This is the logic that is attempted to be applied
in dbsize.c in this patch.
So what do you think of the patch?
On 2020-09-09 12:41 a.m., gkokolatos@pm.me wrote:
> ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> On Tuesday, 8 September 2020 22:26, David Zhang <david.zhang@highgo.ca> wrote:
>
>>
>> I found the function "table_relation_size" is only used by buffer
>> manager for "RELKIND_RELATION", "RELKIND_TOASTVALUE" and
>> "RELKIND_MATVIEW", i.e.
>>
>> case RELKIND_RELATION:
>> case RELKIND_TOASTVALUE:
>> case RELKIND_MATVIEW:
>> {
>> /*
>> * Not every table AM uses BLCKSZ wide fixed size blocks.
>> * Therefore tableam returns the size in bytes - but
>> for the
>> * purpose of this routine, we want the number of blocks.
>> * Therefore divide, rounding up.
>> */
>> uint64 szbytes;
>>
>> szbytes = table_relation_size(relation, forkNum);
>>
>> return (szbytes + (BLCKSZ - 1)) / BLCKSZ;
>> }
>>
>> So using "calculate_relation_size" and "calculate_toast_table_size" in
>> "calculate_table_size" is easy to understand and the original logic is
>> simple.
>>
> You are correct. This is the logic that is attempted to be applied
> in dbsize.c in this patch.
>
> So what do you think of the patch?
I would suggest to keep the original logic unless there is a bug.
--
David
Software Engineer
Highgo Software Inc. (Canada)
www.highgo.ca
>> So what do you think of the patch?
>
> I would suggest to keep the original logic unless there is a bug.
IIUC, the premise of this path submission is that this codepath is in fact
buggy as it may lead to incorrect results for non-heap relations?
Since we have introduced the table AM api I support going throug it for all
table accesses, so +1 on the overall idea.
Some comments on the patch:
- * Note that this also behaves sanely if applied to an index or toast table;
+ * Note that this also behaves sanely if applied to a toast table;
* those won't have attached toast tables, but they can have multiple forks.
This comment reads a bit odd now and should probably be reworded.
- return size;
+ Assert(size < PG_INT64_MAX);
+
+ return (int64)size;
I assume that this change, and the other ones like that, aim to handle int64
overflow? Using the extra legroom of uint64 can still lead to an overflow,
however theoretical it may be. Wouldn't it be better to check for overflow
before adding to size rather than after the fact? Something along the lines of
checking for headroom left:
rel_size = table_relation_size(..);
if (rel_size > (PG_INT64_MAX - total_size))
< error codepath >
total_size += rel_size;
+ if (rel->rd_rel->relkind != RELKIND_INDEX)
+ {
+ relation_close(rel, AccessShareLock);
+ PG_RETURN_NULL();
+ }
pg_indexes_size is defined as returning the size of the indexes attached to the
specified relation, so this hunk is wrong as it instead requires rel to be an
index?
cheers ./daniel
On Thu, Sep 10, 2020 at 11:51:30AM +0200, Daniel Gustafsson wrote: > Some comments on the patch: Extra comment for this patch: regression tests are failing. -- Michael
Attachment
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Thursday, September 10, 2020 12:51 PM, Daniel Gustafsson <daniel@yesql.se> wrote:
[snip]
> Since we have introduced the table AM api I support going throug it for all
> table accesses, so +1 on the overall idea.
>
Thank you for reviewing! Please find v2 of the patch attached.
In addition to addressing the comments, this patch contains a slightly opinionated approach during describe. In short,
onlyrelations that have storage, are returning non null size during when \d*+ commands are emitted. Such an example is
aview which can be found in the psql tests. If a view was returning a size of 0 Bytes, it would indicate that it can
potentiallybe non zero, which is of course wrong.
> Some comments on the patch:
>
> - - Note that this also behaves sanely if applied to an index or toast table;
>
> - - Note that this also behaves sanely if applied to a toast table;
> - those won't have attached toast tables, but they can have multiple forks.
> This comment reads a bit odd now and should probably be reworded.
>
Agreed and amended.
>
> - return size;
>
> - Assert(size < PG_INT64_MAX);
>
> -
> - return (int64)size;
> I assume that this change, and the other ones like that, aim to handle int64
> overflow? Using the extra legroom of uint64 can still lead to an overflow,
> however theoretical it may be. Wouldn't it be better to check for overflow
> before adding to size rather than after the fact? Something along the lines of
> checking for headroom left:
>
> rel_size = table_relation_size(..);
> if (rel_size > (PG_INT64_MAX - total_size))
> < error codepath >
>
>
> total_size += rel_size;
Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the calling
functionreturns int64.
The Assert() has been placed in order to denote that it is acknowledged that the two functions return different types.
Iwas of the opinion that a run time check will not be needed as even the smaller type can cover more than 9200 PetaByte
tables.
If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functions to
returnuint64 instead.
>
> - if (rel->rd_rel->relkind != RELKIND_INDEX)
>
> - {
>
> - relation_close(rel, AccessShareLock);
>
>
> - PG_RETURN_NULL();
>
>
> - }
> pg_indexes_size is defined as returning the size of the indexes attached to the
> specified relation, so this hunk is wrong as it instead requires rel to be an
> index?
You are absolutely correct, amended.
>
> cheers ./daniel
>
Attachment
Hey Georgios,
Thanks for looking for more avenues to invoke tableAM APIS! Please find
my review below:
On Tue, Oct 13, 2020 at 6:28 AM <gkokolatos@pm.me> wrote:
1.
> /*
> - * heap size, including FSM and VM
> + * table size, including FSM and VM
> */
We should not mention FSM and VM in dbsize.c at all as these are
heap AM specific. We can say:
table size, excluding TOAST relation
2.
> /*
> * Size of toast relation
> */
> if (OidIsValid(rel->rd_rel->reltoastrelid))
> - size += calculate_toast_table_size(rel->rd_rel->reltoastrelid);
> + {
> + Relation toastRel;
> +
> + toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock);
We can replace the OidIsValid check with relation_needs_toast_table()
and then have the OidIsValid() check in an Assert. Perhaps, that will
make things more readable.
3.
> + if (rel->rd_rel->relkind == RELKIND_RELATION ||
> + rel->rd_rel->relkind == RELKIND_TOASTVALUE ||
> + rel->rd_rel->relkind == RELKIND_MATVIEW)
> + size = calculate_table_size(rel);
> + else
> + {
> + relation_close(rel, AccessShareLock);
> + PG_RETURN_NULL();
> + }
This leads to behavioral changes:
I am talking about the case where one calls pg_table_size() on an index.
W/o your patch, it returns the size of the index. W/ your patch it
returns NULL. Judging by the documentation, this function should not
ideally apply to indexes but it does! I have a sinking feeling that lots
of users use it for this purpose, as there is no function to calculate
the size of a single specific index (except pg_relation_size()).
The same argument I have made above applies to sequences. Users may have
trial-and-errored their way into finding that pg_table_size() can tell
them the size of a specific index/sequence! I don't know how widespread
the use is in the user community, so IMO maybe we should be conservative
and not introduce this change? Alternatively, we could call out that
pg_table_size() is only for tables by throwing an error if anything
other than a table is passed in.
If we decide to preserve the existing behavior of the pg_table_size():
It seems that for things not backed by the tableAM (indexes and
sequences), they should still go through calculate_relation_size().
We can call table_relation_size() based on if relkind is
RELKIND_RELATION, RELKIND_TOASTVALUE or RELKIND_MATVIEW. Perhaps it
might be worthwhile making a new macro RELKIND_HAS_TABLE_STORAGE to
capture these three cases (See RELKIND_HAS_STORAGE). This also ensures
that we return 0 for things that don't qualify as RELKIND_HAS_STORAGE,
such as a partitioned table (Currently w/ the patch applied, we return
NULL for those cases, which is another behavior change)
4.
> @@ -3776,10 +3776,24 @@ listTables(const char *tabtypes, const char *pattern, bool verbose, bool showSys
> gettext_noop("Access Method"));
>
> /*
> + * As of PostgreSQL 14, do not use pg_table_size() for indexes and
> + * sequences as it does not behave sanely for those.
> + *
> * As of PostgreSQL 9.0, use pg_table_size() to show a more accurate
> * size of a table, including FSM, VM and TOAST tables.
> */
> - if (pset.sversion >= 90000)
> + if (pset.sversion >= 140000)
> + appendPQExpBuffer(&buf,
> + ",\n CASE"
> + " WHEN c.relkind in ("CppAsString2(RELKIND_INDEX)", "
> + CppAsString2(RELKIND_PARTITIONED_INDEX)", "
> + CppAsString2(RELKIND_SEQUENCE)") THEN"
> + " pg_catalog.pg_size_pretty(pg_catalog.pg_relation_size(c.oid))"
> + " ELSE"
> + " pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid))"
> + " END as \"%s\"",
> + gettext_noop("Size"));
> + else if (pset.sversion >= 90000)
> appendPQExpBuffer(&buf,
> ",\n pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid)) as \"%s\"",
> gettext_noop("Size"));
Following on from point 3, if we decide to preserve the existing behavior,
we wouldn't need this change, as it would be internalized within
pg_table_size().
4.
> >
> > - return size;
> >
> > - Assert(size < PG_INT64_MAX);
> >
> > -
> > - return (int64)size;
> > I assume that this change, and the other ones like that, aim to handle int64
> > overflow? Using the extra legroom of uint64 can still lead to an overflow,
> > however theoretical it may be. Wouldn't it be better to check for overflow
> > before adding to size rather than after the fact? Something along the lines of
> > checking for headroom left:
> >
> > rel_size = table_relation_size(..);
> > if (rel_size > (PG_INT64_MAX - total_size))
> > < error codepath >
> >
> >
> > total_size += rel_size;
>
>
>
> Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the calling
functionreturns int64.
>
>
>
> The Assert() has been placed in order to denote that it is acknowledged that the two functions return different
types.I was of the opinion that a run time check will not be needed as even the > smaller type can cover more than 9200
PetaBytetables.
>
>
>
> If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functions to
returnuint64 instead.
Changing the signature would be the ideal change for all of this here.
But Postgres does not have support for an unsigned 64bit integer (bigint
is signed). One would need to turn to extensions such as [1]. Thus, +1
to what Daniel said above.
5.
> @@ -415,7 +384,7 @@ calculate_table_size(Relation rel)
> static int64
> calculate_indexes_size(Relation rel)
> {
> - int64 size = 0;
> + uint64 size = 0;
>
> /*
> * Aggregate all indexes on the given relation
> @@ -444,7 +413,9 @@ calculate_indexes_size(Relation rel)
> list_free(index_oids);
> }
>
> - return size;
> + Assert(size < PG_INT64_MAX);
> +
> + return (int64)size;
> }
I don't think we would need these changes as nothing changed in this
function.
Regards,
Soumyadeep
[1] https://github.com/petere/pguint
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ On Monday, November 9, 2020 7:50 PM, Soumyadeep Chakraborty <soumyadeep2007@gmail.com> wrote: > Hey Georgios, > > Thanks for looking for more avenues to invoke tableAM APIS! Please find > my review below: A great review Soumyadeep, it is much appreciated. Please remember to add yourself as a reviewer in the commitfest [https://commitfest.postgresql.org/30/2701/] > > On Tue, Oct 13, 2020 at 6:28 AM gkokolatos@pm.me wrote: > > 1. > > > /* > > > > - - heap size, including FSM and VM > > > > - - table size, including FSM and VM > > */ > > > > We should not mention FSM and VM in dbsize.c at all as these are > heap AM specific. We can say: > table size, excluding TOAST relation Yeah, I was thinking that the notion that FSM and VM are still taken into account should be stated. We are iterating over ForkNumber after all. How about phrasing it as: + table size, including all implemented forks from the AM (e.g. FSM, VM) + excluding TOAST relations Thoughts? > > 2. > > > /* > > > > - Size of toast relation > > */ > > if (OidIsValid(rel->rd_rel->reltoastrelid)) > > > > > > - size += calculate_toast_table_size(rel->rd_rel->reltoastrelid); > > > > - { > > - Relation toastRel; > > - > > - toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > We can replace the OidIsValid check with relation_needs_toast_table() > and then have the OidIsValid() check in an Assert. Perhaps, that will > make things more readable. Please, allow me to kindly disagree. Relation is already open at this stage. Even create_toast_table(), the internal workhorse for creating toast relations, does check reltoastrelid to test if the relation is already toasted. Furthermore, we do call: + toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); and in order to avoid elog() errors underneath us, we ought to have verified the validity of reltoastrelid. In short, I think that the code in the proposal is not too unreadable nor that it breaks the coding patterns throughout the codebase. Am I too wrong? > > 3. > > > - if (rel->rd_rel->relkind == RELKIND_RELATION || > > - rel->rd_rel->relkind == RELKIND_TOASTVALUE || > > - rel->rd_rel->relkind == RELKIND_MATVIEW) > > - size = calculate_table_size(rel); > > - else > > - { > > - relation_close(rel, AccessShareLock); > > - PG_RETURN_NULL(); > > - } > > This leads to behavioral changes: > > I am talking about the case where one calls pg_table_size() on an index. > W/o your patch, it returns the size of the index. W/ your patch it > returns NULL. Judging by the documentation, this function should not > ideally apply to indexes but it does! I have a sinking feeling that lots > of users use it for this purpose, as there is no function to calculate > the size of a single specific index (except pg_relation_size()). > The same argument I have made above applies to sequences. Users may have > trial-and-errored their way into finding that pg_table_size() can tell > them the size of a specific index/sequence! I don't know how widespread > the use is in the user community, so IMO maybe we should be conservative > and not introduce this change? Alternatively, we could call out that > pg_table_size() is only for tables by throwing an error if anything > other than a table is passed in. > > If we decide to preserve the existing behavior of the pg_table_size(): > It seems that for things not backed by the tableAM (indexes and > sequences), they should still go through calculate_relation_size(). > We can call table_relation_size() based on if relkind is > RELKIND_RELATION, RELKIND_TOASTVALUE or RELKIND_MATVIEW. Perhaps it > might be worthwhile making a new macro RELKIND_HAS_TABLE_STORAGE to > capture these three cases (See RELKIND_HAS_STORAGE). This also ensures > that we return 0 for things that don't qualify as RELKIND_HAS_STORAGE, > such as a partitioned table (Currently w/ the patch applied, we return > NULL for those cases, which is another behavior change) Excellent point. This is the discussion I was longing to have. I stand by the decision coded in the patch, that pg_table_size() should return NULL for other kinds of relations, such as indexes, sequences etc. It is a conscious decision based on the following: * Supported by the documentation, pg_table_size() applies to tables only. For other uses the higher-level functions pg_total_relation_size() or pg_relation_size() should be used. * Commit fa352d662e taught pg_relation_size() and friends to return NULL if the object doesn't exist. This makes perfectsense for the scenarios described in the commit: That avoids errors when the functions are used in queries like "SELECT pg_relation_size(oid) FROM pg_class", and a table is dropped concurrently. IMHO: It is more consistent to return NULL when the relation does exist OR it is not a table kind. * Returning 0 for things that do not have storage, is nonsensical because it implies that it can be NON zero at some point. Things that do not have storage have an unknown size. As far as for the argument that users might have trialed and errored their way into undocumented behaviour, I do not think it is strong enough to stop us from implementing the documented behaviour. > > 4. > > > @@ -3776,10 +3776,24 @@ listTables(const char *tabtypes, const char *pattern, bool verbose, bool showSys > > gettext_noop("Access Method")); > > > > /* > > > > > > - * As of PostgreSQL 14, do not use pg_table_size() for indexes and > > > > > > - * sequences as it does not behave sanely for those. > > > > > > - * > > * As of PostgreSQL 9.0, use pg_table_size() to show a more accurate > > * size of a table, including FSM, VM and TOAST tables. > > */ > > > > > > > > - if (pset.sversion >= 90000) > > > > > > > > - if (pset.sversion >= 140000) > > > > > > - appendPQExpBuffer(&buf, > > > > > > - ",\\n CASE" > > > > > > - " WHEN c.relkind in ("CppAsString2(RELKIND_INDEX)", " > > > > > > - CppAsString2(RELKIND_PARTITIONED_INDEX)", " > > > > > > - CppAsString2(RELKIND_SEQUENCE)") THEN" > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_relation_size(c.oid))" > > > > > > - " ELSE" > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid))" > > > > > > - " END as \\"%s\\"", > > > > > > - gettext_noop("Size")); > > > > > > - else if (pset.sversion >= 90000) > > appendPQExpBuffer(&buf, > > ",\\n pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid)) as \\"%s\\"", > > gettext_noop("Size")); > > > > > > Following on from point 3, if we decide to preserve the existing behavior, > we wouldn't need this change, as it would be internalized within > pg_table_size(). We really should not decide to preserve the existing behaviour. I will reiterate my point: Returning 0 for things that do not have storage, implies that it can be NON zero at some point. We should not treat pg_table_size() as an alias for pg_relation_size(). > > 4. > > > > - return size; > > > > > > - Assert(size < PG_INT64_MAX); > > > > > > - > > > - return (int64)size; > > > I assume that this change, and the other ones like that, aim to handle int64 > > > overflow? Using the extra legroom of uint64 can still lead to an overflow, > > > however theoretical it may be. Wouldn't it be better to check for overflow > > > before adding to size rather than after the fact? Something along the lines of > > > checking for headroom left: > > > rel_size = table_relation_size(..); > > > if (rel_size > (PG_INT64_MAX - total_size)) > > > < error codepath > > > > > > > > > > total_size += rel_size; > > > > Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the calling functionreturns int64. > > The Assert() has been placed in order to denote that it is acknowledged that the two functions return different types.I was of the opinion that a run time check will not be needed as even the > smaller type can cover more than 9200 PetaBytetables. > > If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functions to returnuint64 instead. > > Changing the signature would be the ideal change for all of this here. > But Postgres does not have support for an unsigned 64bit integer (bigint > is signed). One would need to turn to extensions such as [1]. Thus, +1 > to what Daniel said above. Apologies, I do not follow. Are you suggesting that we should introduce overflow tests? > > 5. > > > @@ -415,7 +384,7 @@ calculate_table_size(Relation rel) > > static int64 > > calculate_indexes_size(Relation rel) > > { > > > > - int64 size = 0; > > > > - uint64 size = 0; > > /* > > - Aggregate all indexes on the given relation > > @@ -444,7 +413,9 @@ calculate_indexes_size(Relation rel) > > list_free(index_oids); > > } > > > > > > - return size; > > > > - Assert(size < PG_INT64_MAX); > > - > > - return (int64)size; > > } > > > > I don't think we would need these changes as nothing changed in this > function. This change intended to keep the calculate family of functions homogenous in style. From the point above I understand it confuses more than helps, I will remove. Cheers, //Georgios > > Regards, > Soumyadeep > > [1] https://github.com/petere/pguint
Hey Georgios, On Tue, Nov 10, 2020 at 6:20 AM <gkokolatos@pm.me> wrote: > > > > > > > ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ > On Monday, November 9, 2020 7:50 PM, Soumyadeep Chakraborty <soumyadeep2007@gmail.com> wrote: > > > Hey Georgios, > > > > Thanks for looking for more avenues to invoke tableAM APIS! Please find > > my review below: > > A great review Soumyadeep, it is much appreciated. > Please remember to add yourself as a reviewer in the commitfest > [https://commitfest.postgresql.org/30/2701/] Ah yes. Sorry, I forgot that! > > > > On Tue, Oct 13, 2020 at 6:28 AM gkokolatos@pm.me wrote: > > > > 1. > > > > > /* > > > > > > - - heap size, including FSM and VM > > > > > > - - table size, including FSM and VM > > > */ > > > > > > > We should not mention FSM and VM in dbsize.c at all as these are > > heap AM specific. We can say: > > table size, excluding TOAST relation > > Yeah, I was thinking that the notion that FSM and VM are still taken > into account should be stated. We are iterating over ForkNumber > after all. > > How about phrasing it as: > > + table size, including all implemented forks from the AM (e.g. FSM, VM) > + excluding TOAST relations > > Thoughts? Yes, I was thinking along the same lines. The concept of a "fork" forced should not be forced down into the tableAM. But that is a discussion for another day. We can probably say: + table size, including all implemented forks from the AM (e.g. FSM, VM + for the heap AM) excluding TOAST relations > > > > 2. > > > > > /* > > > > > > - Size of toast relation > > > */ > > > if (OidIsValid(rel->rd_rel->reltoastrelid)) > > > > > > > > > - size += calculate_toast_table_size(rel->rd_rel->reltoastrelid); > > > > > > - { > > > - Relation toastRel; > > > - > > > - toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > > > We can replace the OidIsValid check with relation_needs_toast_table() > > and then have the OidIsValid() check in an Assert. Perhaps, that will > > make things more readable. > > Please, allow me to kindly disagree. > > Relation is already open at this stage. Even create_toast_table(), the > internal workhorse for creating toast relations, does check reltoastrelid > to test if the relation is already toasted. > > Furthermore, we do call: > > + toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > and in order to avoid elog() errors underneath us, we ought to have > verified the validity of reltoastrelid. > > In short, I think that the code in the proposal is not too unreadable > nor that it breaks the coding patterns throughout the codebase. > > Am I too wrong? No not at all. The code in the proposal is indeed adhering to the codebase. What I was going for here was to increase the usage of relation_needs_toast_table(). What I meant was: if (table_relation_needs_toast_table(rel)) { if (!OidIsValid(rel->rd_rel->reltoastrelid)) { elog(ERROR, <errmsg that toast table wasn't found>); } //size calculation here.. } We want to error out if the toast table can't be found and the relation is expected to have one, which the existing code doesn't handle. > > > > > 3. > > > > > - if (rel->rd_rel->relkind == RELKIND_RELATION || > > > - rel->rd_rel->relkind == RELKIND_TOASTVALUE || > > > - rel->rd_rel->relkind == RELKIND_MATVIEW) > > > - size = calculate_table_size(rel); > > > - else > > > - { > > > - relation_close(rel, AccessShareLock); > > > - PG_RETURN_NULL(); > > > - } > > > > This leads to behavioral changes: > > > > I am talking about the case where one calls pg_table_size() on an index. > > W/o your patch, it returns the size of the index. W/ your patch it > > returns NULL. Judging by the documentation, this function should not > > ideally apply to indexes but it does! I have a sinking feeling that lots > > of users use it for this purpose, as there is no function to calculate > > the size of a single specific index (except pg_relation_size()). > > The same argument I have made above applies to sequences. Users may have > > trial-and-errored their way into finding that pg_table_size() can tell > > them the size of a specific index/sequence! I don't know how widespread > > the use is in the user community, so IMO maybe we should be conservative > > and not introduce this change? Alternatively, we could call out that > > pg_table_size() is only for tables by throwing an error if anything > > other than a table is passed in. > > > > If we decide to preserve the existing behavior of the pg_table_size(): > > It seems that for things not backed by the tableAM (indexes and > > sequences), they should still go through calculate_relation_size(). > > We can call table_relation_size() based on if relkind is > > RELKIND_RELATION, RELKIND_TOASTVALUE or RELKIND_MATVIEW. Perhaps it > > might be worthwhile making a new macro RELKIND_HAS_TABLE_STORAGE to > > capture these three cases (See RELKIND_HAS_STORAGE). This also ensures > > that we return 0 for things that don't qualify as RELKIND_HAS_STORAGE, > > such as a partitioned table (Currently w/ the patch applied, we return > > NULL for those cases, which is another behavior change) > > > Excellent point. This is the discussion I was longing to have. > > I stand by the decision coded in the patch, that pg_table_size() should > return NULL for other kinds of relations, such as indexes, sequences > etc. > > It is a conscious decision based on the following: > > * Supported by the documentation, pg_table_size() applies to tables only. > For other uses the higher-level functions pg_total_relation_size() or > pg_relation_size() should be used. > * Commit fa352d662e taught pg_relation_size() and friends to return NULL if the object doesn't exist. This makes perfectsense for the > scenarios described in the commit: > > That avoids errors when the functions are used in queries like > "SELECT pg_relation_size(oid) FROM pg_class", > and a table is dropped concurrently. > > IMHO: It is more consistent to return NULL when the relation does exist > OR it is not a table kind. > * Returning 0 for things that do not have storage, is nonsensical because > it implies that it can be NON zero at some point. Things that do not > have storage have an unknown size. Fair. We will have to document the behavior change. > > As far as for the argument that users might have trialed and errored > their way into undocumented behaviour, I do not think it is strong > enough to stop us from implementing the documented behaviour. Fair. I would strongly vote for having two additional functions (pg_index_size() and pg_sequence_size()) to strongly signal our intent of banning that kind of use of pg_table_size(). I think it would help users a lot. It is not easy to find what function to call when you want the size of a single index/sequence. > > > > 4. > > > > > @@ -3776,10 +3776,24 @@ listTables(const char *tabtypes, const char *pattern, bool verbose, bool showSys > > > gettext_noop("Access Method")); > > > > > > /* > > > > > > > > > - * As of PostgreSQL 14, do not use pg_table_size() for indexes and > > > > > > > > > - * sequences as it does not behave sanely for those. > > > > > > > > > - * > > > * As of PostgreSQL 9.0, use pg_table_size() to show a more accurate > > > * size of a table, including FSM, VM and TOAST tables. > > > */ > > > > > > > > > > > > - if (pset.sversion >= 90000) > > > > > > > > > > > > - if (pset.sversion >= 140000) > > > > > > > > > - appendPQExpBuffer(&buf, > > > > > > > > > - ",\\n CASE" > > > > > > > > > - " WHEN c.relkind in ("CppAsString2(RELKIND_INDEX)", " > > > > > > > > > - CppAsString2(RELKIND_PARTITIONED_INDEX)", " > > > > > > > > > - CppAsString2(RELKIND_SEQUENCE)") THEN" > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_relation_size(c.oid))" > > > > > > > > > - " ELSE" > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid))" > > > > > > > > > - " END as \\"%s\\"", > > > > > > > > > - gettext_noop("Size")); > > > > > > > > > - else if (pset.sversion >= 90000) > > > appendPQExpBuffer(&buf, > > > ",\\n pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid)) as \\"%s\\"", > > > gettext_noop("Size")); > > > > > > > > > > Following on from point 3, if we decide to preserve the existing behavior, > > we wouldn't need this change, as it would be internalized within > > pg_table_size(). > > We really should not decide to preserve the existing behaviour. > > I will reiterate my point: Returning 0 for things that do not have > storage, implies that it can be NON zero at some point. We should not > treat pg_table_size() as an alias for pg_relation_size(). +1 > > > > 4. > > > > > > - return size; > > > > > > > > - Assert(size < PG_INT64_MAX); > > > > > > > > - > > > > - return (int64)size; > > > > I assume that this change, and the other ones like that, aim to handle int64 > > > > overflow? Using the extra legroom of uint64 can still lead to an overflow, > > > > however theoretical it may be. Wouldn't it be better to check for overflow > > > > before adding to size rather than after the fact? Something along the lines of > > > > checking for headroom left: > > > > rel_size = table_relation_size(..); > > > > if (rel_size > (PG_INT64_MAX - total_size)) > > > > < error codepath > > > > > > > > > > > > > total_size += rel_size; > > > > > > Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the calling functionreturns int64. > > > The Assert() has been placed in order to denote that it is acknowledged that the two functions return different types.I was of the opinion that a run time check will not be needed as even the > smaller type can cover more than 9200 PetaBytetables. > > > If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functions toreturn uint64 instead. > > > > Changing the signature would be the ideal change for all of this here. > > But Postgres does not have support for an unsigned 64bit integer (bigint > > is signed). One would need to turn to extensions such as [1]. Thus, +1 > > to what Daniel said above. > > Apologies, I do not follow. Are you suggesting that we should > introduce overflow tests? Yes, to introduce the same overflow test that Daniel suggested above as returning a uint64 in PG does not really return a uint64 AFAIU (since the pg_**size() functions all return bigint which is signed and there is no uint64 user-facing type). Regards, Soumyadeep (VMware)
On Thu, Nov 12, 2020 at 2:54 AM Soumyadeep Chakraborty <soumyadeep2007@gmail.com> wrote: > > Hey Georgios, > > On Tue, Nov 10, 2020 at 6:20 AM <gkokolatos@pm.me> wrote: > > > > > > > > > > > > > > ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ > > On Monday, November 9, 2020 7:50 PM, Soumyadeep Chakraborty <soumyadeep2007@gmail.com> wrote: > > > > > Hey Georgios, > > > > > > Thanks for looking for more avenues to invoke tableAM APIS! Please find > > > my review below: > > > > A great review Soumyadeep, it is much appreciated. > > Please remember to add yourself as a reviewer in the commitfest > > [https://commitfest.postgresql.org/30/2701/] > > Ah yes. Sorry, I forgot that! > > > > > > > On Tue, Oct 13, 2020 at 6:28 AM gkokolatos@pm.me wrote: > > > > > > 1. > > > > > > > /* > > > > > > > > - - heap size, including FSM and VM > > > > > > > > - - table size, including FSM and VM > > > > */ > > > > > > > > > > We should not mention FSM and VM in dbsize.c at all as these are > > > heap AM specific. We can say: > > > table size, excluding TOAST relation > > > > Yeah, I was thinking that the notion that FSM and VM are still taken > > into account should be stated. We are iterating over ForkNumber > > after all. > > > > How about phrasing it as: > > > > + table size, including all implemented forks from the AM (e.g. FSM, VM) > > + excluding TOAST relations > > > > Thoughts? > > Yes, I was thinking along the same lines. The concept of a "fork" forced > should not be forced down into the tableAM. But that is a discussion for > another day. We can probably say: > > + table size, including all implemented forks from the AM (e.g. FSM, VM > + for the heap AM) excluding TOAST relations > > > > > > > 2. > > > > > > > /* > > > > > > > > - Size of toast relation > > > > */ > > > > if (OidIsValid(rel->rd_rel->reltoastrelid)) > > > > > > > > > > > > - size += calculate_toast_table_size(rel->rd_rel->reltoastrelid); > > > > > > > > - { > > > > - Relation toastRel; > > > > - > > > > - toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > > > > > We can replace the OidIsValid check with relation_needs_toast_table() > > > and then have the OidIsValid() check in an Assert. Perhaps, that will > > > make things more readable. > > > > Please, allow me to kindly disagree. > > > > Relation is already open at this stage. Even create_toast_table(), the > > internal workhorse for creating toast relations, does check reltoastrelid > > to test if the relation is already toasted. > > > > Furthermore, we do call: > > > > + toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > > > and in order to avoid elog() errors underneath us, we ought to have > > verified the validity of reltoastrelid. > > > > In short, I think that the code in the proposal is not too unreadable > > nor that it breaks the coding patterns throughout the codebase. > > > > Am I too wrong? > > No not at all. The code in the proposal is indeed adhering to the > codebase. What I was going for here was to increase the usage of > relation_needs_toast_table(). What I meant was: > > if (table_relation_needs_toast_table(rel)) > { > if (!OidIsValid(rel->rd_rel->reltoastrelid)) > { > elog(ERROR, <errmsg that toast table wasn't found>); > } > //size calculation here.. > } > > We want to error out if the toast table can't be found and the relation > is expected to have one, which the existing code doesn't handle. > > > > > > > > > > 3. > > > > > > > - if (rel->rd_rel->relkind == RELKIND_RELATION || > > > > - rel->rd_rel->relkind == RELKIND_TOASTVALUE || > > > > - rel->rd_rel->relkind == RELKIND_MATVIEW) > > > > - size = calculate_table_size(rel); > > > > - else > > > > - { > > > > - relation_close(rel, AccessShareLock); > > > > - PG_RETURN_NULL(); > > > > - } > > > > > > This leads to behavioral changes: > > > > > > I am talking about the case where one calls pg_table_size() on an index. > > > W/o your patch, it returns the size of the index. W/ your patch it > > > returns NULL. Judging by the documentation, this function should not > > > ideally apply to indexes but it does! I have a sinking feeling that lots > > > of users use it for this purpose, as there is no function to calculate > > > the size of a single specific index (except pg_relation_size()). > > > The same argument I have made above applies to sequences. Users may have > > > trial-and-errored their way into finding that pg_table_size() can tell > > > them the size of a specific index/sequence! I don't know how widespread > > > the use is in the user community, so IMO maybe we should be conservative > > > and not introduce this change? Alternatively, we could call out that > > > pg_table_size() is only for tables by throwing an error if anything > > > other than a table is passed in. > > > > > > If we decide to preserve the existing behavior of the pg_table_size(): > > > It seems that for things not backed by the tableAM (indexes and > > > sequences), they should still go through calculate_relation_size(). > > > We can call table_relation_size() based on if relkind is > > > RELKIND_RELATION, RELKIND_TOASTVALUE or RELKIND_MATVIEW. Perhaps it > > > might be worthwhile making a new macro RELKIND_HAS_TABLE_STORAGE to > > > capture these three cases (See RELKIND_HAS_STORAGE). This also ensures > > > that we return 0 for things that don't qualify as RELKIND_HAS_STORAGE, > > > such as a partitioned table (Currently w/ the patch applied, we return > > > NULL for those cases, which is another behavior change) > > > > > > Excellent point. This is the discussion I was longing to have. > > > > I stand by the decision coded in the patch, that pg_table_size() should > > return NULL for other kinds of relations, such as indexes, sequences > > etc. > > > > It is a conscious decision based on the following: > > > > * Supported by the documentation, pg_table_size() applies to tables only. > > For other uses the higher-level functions pg_total_relation_size() or > > pg_relation_size() should be used. > > * Commit fa352d662e taught pg_relation_size() and friends to return NULL if the object doesn't exist. This makes perfectsense for the > > scenarios described in the commit: > > > > That avoids errors when the functions are used in queries like > > "SELECT pg_relation_size(oid) FROM pg_class", > > and a table is dropped concurrently. > > > > IMHO: It is more consistent to return NULL when the relation does exist > > OR it is not a table kind. > > * Returning 0 for things that do not have storage, is nonsensical because > > it implies that it can be NON zero at some point. Things that do not > > have storage have an unknown size. > > Fair. We will have to document the behavior change. > > > > > As far as for the argument that users might have trialed and errored > > their way into undocumented behaviour, I do not think it is strong > > enough to stop us from implementing the documented behaviour. > > Fair. I would strongly vote for having two additional functions > (pg_index_size() and pg_sequence_size()) to strongly signal our intent > of banning that kind of use of pg_table_size(). I think it would help > users a lot. It is not easy to find what function to call when you want > the size of a single index/sequence. > > > > > > > 4. > > > > > > > @@ -3776,10 +3776,24 @@ listTables(const char *tabtypes, const char *pattern, bool verbose, bool showSys > > > > gettext_noop("Access Method")); > > > > > > > > /* > > > > > > > > > > > > - * As of PostgreSQL 14, do not use pg_table_size() for indexes and > > > > > > > > > > > > - * sequences as it does not behave sanely for those. > > > > > > > > > > > > - * > > > > * As of PostgreSQL 9.0, use pg_table_size() to show a more accurate > > > > * size of a table, including FSM, VM and TOAST tables. > > > > */ > > > > > > > > > > > > > > > > - if (pset.sversion >= 90000) > > > > > > > > > > > > > > > > - if (pset.sversion >= 140000) > > > > > > > > > > > > - appendPQExpBuffer(&buf, > > > > > > > > > > > > - ",\\n CASE" > > > > > > > > > > > > - " WHEN c.relkind in ("CppAsString2(RELKIND_INDEX)", " > > > > > > > > > > > > - CppAsString2(RELKIND_PARTITIONED_INDEX)", " > > > > > > > > > > > > - CppAsString2(RELKIND_SEQUENCE)") THEN" > > > > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_relation_size(c.oid))" > > > > > > > > > > > > - " ELSE" > > > > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid))" > > > > > > > > > > > > - " END as \\"%s\\"", > > > > > > > > > > > > - gettext_noop("Size")); > > > > > > > > > > > > - else if (pset.sversion >= 90000) > > > > appendPQExpBuffer(&buf, > > > > ",\\n pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid)) as \\"%s\\"", > > > > gettext_noop("Size")); > > > > > > > > > > > > > > Following on from point 3, if we decide to preserve the existing behavior, > > > we wouldn't need this change, as it would be internalized within > > > pg_table_size(). > > > > We really should not decide to preserve the existing behaviour. > > > > I will reiterate my point: Returning 0 for things that do not have > > storage, implies that it can be NON zero at some point. We should not > > treat pg_table_size() as an alias for pg_relation_size(). > > +1 > > > > > > > 4. > > > > > > > > - return size; > > > > > > > > > > - Assert(size < PG_INT64_MAX); > > > > > > > > > > - > > > > > - return (int64)size; > > > > > I assume that this change, and the other ones like that, aim to handle int64 > > > > > overflow? Using the extra legroom of uint64 can still lead to an overflow, > > > > > however theoretical it may be. Wouldn't it be better to check for overflow > > > > > before adding to size rather than after the fact? Something along the lines of > > > > > checking for headroom left: > > > > > rel_size = table_relation_size(..); > > > > > if (rel_size > (PG_INT64_MAX - total_size)) > > > > > < error codepath > > > > > > > > > > > > > > > > total_size += rel_size; > > > > > > > > Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the callingfunction returns int64. > > > > The Assert() has been placed in order to denote that it is acknowledged that the two functions return different types.I was of the opinion that a run time check will not be needed as even the > smaller type can cover more than 9200 PetaBytetables. > > > > If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functions toreturn uint64 instead. > > > > > > Changing the signature would be the ideal change for all of this here. > > > But Postgres does not have support for an unsigned 64bit integer (bigint > > > is signed). One would need to turn to extensions such as [1]. Thus, +1 > > > to what Daniel said above. > > > > Apologies, I do not follow. Are you suggesting that we should > > introduce overflow tests? > > Yes, to introduce the same overflow test that Daniel suggested above as > returning a uint64 in PG does not really return a uint64 AFAIU (since > the pg_**size() functions all return bigint which is signed and there > is no uint64 user-facing type). Status update for a commitfest entry. This patch gets review comments and there was some discussion. It seems we're waiting for the patch update. So I've moved this patch to the next commitfest and set it to "Waiting on Author". Regards, -- Masahiko Sawada EDB: https://www.enterprisedb.com/
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ On Monday, February 1, 2021 1:18 PM, Masahiko Sawada <sawada.mshk@gmail.com> wrote: > On Thu, Nov 12, 2020 at 2:54 AM Soumyadeep Chakraborty > soumyadeep2007@gmail.com wrote: > > > Hey Georgios, > > On Tue, Nov 10, 2020 at 6:20 AM gkokolatos@pm.me wrote: > > > > > ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ > > > On Monday, November 9, 2020 7:50 PM, Soumyadeep Chakraborty soumyadeep2007@gmail.com wrote: > > > > > > > Hey Georgios, > > > > Thanks for looking for more avenues to invoke tableAM APIS! Please find > > > > my review below: > > > > > > A great review Soumyadeep, it is much appreciated. > > > Please remember to add yourself as a reviewer in the commitfest > > > [https://commitfest.postgresql.org/30/2701/] > > > > Ah yes. Sorry, I forgot that! > > > > > > On Tue, Oct 13, 2020 at 6:28 AM gkokolatos@pm.me wrote: > > > > > > > > 1. > > > > > > > > > /* > > > > > > > > > > - - heap size, including FSM and VM > > > > > - - table size, including FSM and VM > > > > > */ > > > > > > > > > > > > > We should not mention FSM and VM in dbsize.c at all as these are > > > > heap AM specific. We can say: > > > > table size, excluding TOAST relation > > > > > > Yeah, I was thinking that the notion that FSM and VM are still taken > > > into account should be stated. We are iterating over ForkNumber > > > after all. > > > How about phrasing it as: > > > > > > - table size, including all implemented forks from the AM (e.g. FSM, VM) > > > - excluding TOAST relations > > > > > > Thoughts? > > > > Yes, I was thinking along the same lines. The concept of a "fork" forced > > should not be forced down into the tableAM. But that is a discussion for > > another day. We can probably say: > > > > - table size, including all implemented forks from the AM (e.g. FSM, VM > > - for the heap AM) excluding TOAST relations > > > > > > 2. > > > > > > > > > /* > > > > > > > > > > - Size of toast relation > > > > > */ > > > > > if (OidIsValid(rel->rd_rel->reltoastrelid)) > > > > > > > > > > - size += calculate_toast_table_size(rel->rd_rel->reltoastrelid); > > > > > > > > > > - { > > > > > > > > > > - Relation toastRel; > > > > > > > > > > - > > > > > - toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > > > > > > > > > > > > We can replace the OidIsValid check with relation_needs_toast_table() > > > > and then have the OidIsValid() check in an Assert. Perhaps, that will > > > > make things more readable. > > > > > > Please, allow me to kindly disagree. > > > Relation is already open at this stage. Even create_toast_table(), the > > > internal workhorse for creating toast relations, does check reltoastrelid > > > to test if the relation is already toasted. > > > Furthermore, we do call: > > > > > > - toastRel = relation_open(rel->rd_rel->reltoastrelid, AccessShareLock); > > > > > > and in order to avoid elog() errors underneath us, we ought to have > > > verified the validity of reltoastrelid. > > > In short, I think that the code in the proposal is not too unreadable > > > nor that it breaks the coding patterns throughout the codebase. > > > Am I too wrong? > > > > No not at all. The code in the proposal is indeed adhering to the > > codebase. What I was going for here was to increase the usage of > > relation_needs_toast_table(). What I meant was: > > if (table_relation_needs_toast_table(rel)) > > { > > if (!OidIsValid(rel->rd_rel->reltoastrelid)) > > { > > elog(ERROR, <errmsg that toast table wasn't found>); > > } > > //size calculation here.. > > } > > We want to error out if the toast table can't be found and the relation > > is expected to have one, which the existing code doesn't handle. > > > > > > 3. > > > > > > > > > - if (rel->rd_rel->relkind == RELKIND_RELATION || > > > > > - rel->rd_rel->relkind == RELKIND_TOASTVALUE || > > > > > - rel->rd_rel->relkind == RELKIND_MATVIEW) > > > > > - size = calculate_table_size(rel); > > > > > - else > > > > > - { > > > > > - relation_close(rel, AccessShareLock); > > > > > - PG_RETURN_NULL(); > > > > > - } > > > > > > > > This leads to behavioral changes: > > > > I am talking about the case where one calls pg_table_size() on an index. > > > > W/o your patch, it returns the size of the index. W/ your patch it > > > > returns NULL. Judging by the documentation, this function should not > > > > ideally apply to indexes but it does! I have a sinking feeling that lots > > > > of users use it for this purpose, as there is no function to calculate > > > > the size of a single specific index (except pg_relation_size()). > > > > The same argument I have made above applies to sequences. Users may have > > > > trial-and-errored their way into finding that pg_table_size() can tell > > > > them the size of a specific index/sequence! I don't know how widespread > > > > the use is in the user community, so IMO maybe we should be conservative > > > > and not introduce this change? Alternatively, we could call out that > > > > pg_table_size() is only for tables by throwing an error if anything > > > > other than a table is passed in. > > > > If we decide to preserve the existing behavior of the pg_table_size(): > > > > It seems that for things not backed by the tableAM (indexes and > > > > sequences), they should still go through calculate_relation_size(). > > > > We can call table_relation_size() based on if relkind is > > > > RELKIND_RELATION, RELKIND_TOASTVALUE or RELKIND_MATVIEW. Perhaps it > > > > might be worthwhile making a new macro RELKIND_HAS_TABLE_STORAGE to > > > > capture these three cases (See RELKIND_HAS_STORAGE). This also ensures > > > > that we return 0 for things that don't qualify as RELKIND_HAS_STORAGE, > > > > such as a partitioned table (Currently w/ the patch applied, we return > > > > NULL for those cases, which is another behavior change) > > > > > > Excellent point. This is the discussion I was longing to have. > > > I stand by the decision coded in the patch, that pg_table_size() should > > > return NULL for other kinds of relations, such as indexes, sequences > > > etc. > > > It is a conscious decision based on the following: > > > > > > - Supported by the documentation, pg_table_size() applies to tables only. > > > For other uses the higher-level functions pg_total_relation_size() or > > > pg_relation_size() should be used. > > > > > > - Commit fa352d662e taught pg_relation_size() and friends to return NULL if the object doesn't exist. This makesperfect sense for the > > > scenarios described in the commit: > > > That avoids errors when the functions are used in queries like > > > "SELECT pg_relation_size(oid) FROM pg_class", > > > and a table is dropped concurrently. > > > > > > > > > IMHO: It is more consistent to return NULL when the relation does exist > > > OR it is not a table kind. > > > > > > - Returning 0 for things that do not have storage, is nonsensical because > > > it implies that it can be NON zero at some point. Things that do not > > > have storage have an unknown size. > > > > > > > Fair. We will have to document the behavior change. > > > > > As far as for the argument that users might have trialed and errored > > > their way into undocumented behaviour, I do not think it is strong > > > enough to stop us from implementing the documented behaviour. > > > > Fair. I would strongly vote for having two additional functions > > (pg_index_size() and pg_sequence_size()) to strongly signal our intent > > of banning that kind of use of pg_table_size(). I think it would help > > users a lot. It is not easy to find what function to call when you want > > the size of a single index/sequence. > > > > > > 4. > > > > > > > > > @@ -3776,10 +3776,24 @@ listTables(const char *tabtypes, const char *pattern, bool verbose, bool showSys > > > > > gettext_noop("Access Method")); > > > > > > > > > > /* > > > > > > > > > > > > > > > - * As of PostgreSQL 14, do not use pg_table_size() for indexes and > > > > > > > > > > > > > > > - * sequences as it does not behave sanely for those. > > > > > > > > > > > > > > > - * > > > > > * As of PostgreSQL 9.0, use pg_table_size() to show a more accurate > > > > > * size of a table, including FSM, VM and TOAST tables. > > > > > */ > > > > > > > > > > > > > > > - if (pset.sversion >= 90000) > > > > > > > > > > > > > > > - if (pset.sversion >= 140000) > > > > > > > > > > > > > > > - appendPQExpBuffer(&buf, > > > > > > > > > > > > > > > - ",\\\\n CASE" > > > > > > > > > > > > > > > - " WHEN c.relkind in ("CppAsString2(RELKIND_INDEX)", " > > > > > > > > > > > > > > > - CppAsString2(RELKIND_PARTITIONED_INDEX)", " > > > > > > > > > > > > > > > - CppAsString2(RELKIND_SEQUENCE)") THEN" > > > > > > > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_relation_size(c.oid))" > > > > > > > > > > > > > > > - " ELSE" > > > > > > > > > > > > > > > - " pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid))" > > > > > > > > > > > > > > > - " END as \\\\"%s\\\\"", > > > > > > > > > > > > > > > - gettext_noop("Size")); > > > > > > > > > > > > > > > - else if (pset.sversion >= 90000) > > > > > appendPQExpBuffer(&buf, > > > > > ",\\\\n pg_catalog.pg_size_pretty(pg_catalog.pg_table_size(c.oid)) as \\\\"%s\\\\"", > > > > > gettext_noop("Size")); > > > > > > > > > > > > > > > > > > Following on from point 3, if we decide to preserve the existing behavior, > > > > we wouldn't need this change, as it would be internalized within > > > > pg_table_size(). > > > > > > We really should not decide to preserve the existing behaviour. > > > I will reiterate my point: Returning 0 for things that do not have > > > storage, implies that it can be NON zero at some point. We should not > > > treat pg_table_size() as an alias for pg_relation_size(). > > > > +1 > > > > > > 4. > > > > > > > > > > - return size; > > > > > > > > > > > > - Assert(size < PG_INT64_MAX); > > > > > > > > > > > > - > > > > > > - return (int64)size; > > > > > > I assume that this change, and the other ones like that, aim to handle int64 > > > > > > overflow? Using the extra legroom of uint64 can still lead to an overflow, > > > > > > however theoretical it may be. Wouldn't it be better to check for overflow > > > > > > before adding to size rather than after the fact? Something along the lines of > > > > > > checking for headroom left: > > > > > > rel_size = table_relation_size(..); > > > > > > if (rel_size > (PG_INT64_MAX - total_size)) > > > > > > < error codepath > > > > > > > > > > > > > > > > > > > total_size += rel_size; > > > > > > > > > > Actually not, the intention is not to handle overflow. The table_relation_size() returns a uint64 and the callingfunction returns int64. > > > > > The Assert() has been placed in order to denote that it is acknowledged that the two functions return differenttypes. I was of the opinion that a run time check will not be needed as even the > smaller type can cover more than9200 PetaByte tables. > > > > > If we were to change anything, then I would prefer to change the signature of the pg_*_size family of functionsto return uint64 instead. > > > > > > > > Changing the signature would be the ideal change for all of this here. > > > > But Postgres does not have support for an unsigned 64bit integer (bigint > > > > is signed). One would need to turn to extensions such as [1]. Thus, +1 > > > > to what Daniel said above. > > > > > > Apologies, I do not follow. Are you suggesting that we should > > > introduce overflow tests? > > > > Yes, to introduce the same overflow test that Daniel suggested above as > > returning a uint64 in PG does not really return a uint64 AFAIU (since > > the pg_**size() functions all return bigint which is signed and there > > is no uint64 user-facing type). > > Status update for a commitfest entry. > > This patch gets review comments and there was some discussion. It > seems we're waiting for the patch update. So I've moved this patch to > the next commitfest and set it to "Waiting on Author". Thank you for your patience. Please find v3 attached. I will reset the status to 'Ready for review'. > > Regards, > > -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- > > Masahiko Sawada > EDB: https://www.enterprisedb.com/
Attachment
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐ On Wednesday, March 17, 2021 6:35 AM, Michael Paquier <michael@paquier.xyz> wrote: > On Mon, Mar 15, 2021 at 03:10:59PM +0900, Michael Paquier wrote: > > > Anyway, as mentioned by other people upthread, I am not really > > convinced either that we should have more flavors of size functions, > > particularly depending on the relkind as this would be confusing for > > the end-user. pg_relation_size() can already do this job for all > > relkinds that use segment files, so it should still be able to hold > > its ground and maintain a consistent behavior with what it does > > currently. > > On top of the rest of my notes, there are two more things that would > face a behavior change if making the existing functions go through > table AMs, which would scan the data in the smgr: I am not certain if you are referring to v5 (sent earlier than your mail) or v4. Can you please verify? > > - After a VACUUM, the relation would be reported with a size of 0, > while that may not be the case of on-disk files yet. I am not really following. Apologies for it. The table AM may or may not choose to go through smgr, depending on the implementation. The only currently known implementation, heap, does invalidate smgr, based on what I can see, after a VACUUM. I have not been able to create a test case both with or without v5 of the patch where not the same result would be returned. What have I missed? > > - Temporary tables of other sessions would be accessible. I am not really certain I am following. Commit 6919b7e3294 from 2012 notes that calculate_relation_size can be safely applied to temp tables of other sessions. v5 of the patch does not change that behaviour. Nor did previous versions, but those are already obsolete. > > So we would likely want a separate function. Another possibility, > which I find tempting, would be to push down the calculation logic > relying on physical files down to the table AM themselves with a new > dedicated callback (relation_size_physical?), as it seems to me that > the most important thing users want to know with this set of functions > is how much physical space is being consumed at one given point in > time. Attached is a small prototype I was just playing with. > -- > Michael >
On Wed, Mar 17, 2021 at 02:35:28PM +0900, Michael Paquier wrote: > So we would likely want a separate function. Another possibility, > which I find tempting, would be to push down the calculation logic > relying on physical files down to the table AM themselves with a new > dedicated callback (relation_size_physical?), as it seems to me that > the most important thing users want to know with this set of functions > is how much physical space is being consumed at one given point in > time. Attached is a small prototype I was just playing with. Thinking more about that, I'd be rather in favor of having a new table AM callback to let the AM measure the size of physical files, and make the business of dbsize.c fall under that. It is clear that this needs more work, so I have marked it as returned with feedback. -- Michael