Thread: Slot's restart_lsn may point to removed WAL segment after hard restart unexpectedly

Dear Hackers,

I'd like to discuss a problem with replication slots's restart LSN. Physical slots are saved to disk at the beginning of checkpoint. At the end of checkpoint, old WAL segments are recycled or removed from disk, if they are not kept by slot's restart_lsn values.

If an existing physical slot is advanced in the middle of checkpoint execution, WAL segments, which are related to saved on disk restart LSN may be removed. It is because the calculation of the replication slot miminal LSN is occured at the end of checkpoint, prior to old WAL segments removal. If to hard stop (pg_stl -m immediate) the postgres instance right after checkpoint and to restart it, the slot's restart_lsn may point to the removed WAL segment. I believe, such behaviour is not good.

The doc [0] describes that restart_lsn may be set to the some past value after reload. There is a discussion [1] on pghackers where such behaviour is discussed. The main reason of not flushing physical slots on advancing is a performance reason. I'm ok with such behaviour, except of that the corresponding WAL segments should not be removed.

I propose to keep WAL segments by saved on disk (flushed) restart_lsn of slots. Add a new field restart_lsn_flushed into ReplicationSlot structure. Copy restart_lsn to restart_lsn_flushed in SaveSlotToPath. It doesn't change the format of storing the slot contents on disk. I attached a patch. It is not yet complete, but demonstate a way to solve the problem.

I reproduced the problem by the following way:

• Add some delay in CheckPointBuffers (pg_usleep) to emulate long checkpoint execution.
• Execute checkpoint and pg_replication_slot_advance right after starting of the checkpoint from another connection.
• Hard restart the server right after checkpoint completion.
• After restart slot's restart_lsn may point to removed WAL segment.

The proposed patch fixes it.

[0] https://www.postgresql.org/docs/current/logicaldecoding-explanation.html
[1] https://www.postgresql.org/message-id/flat/059cc53a-8b14-653a-a24d-5f867503b0ee%40postgrespro.ru

Dear Hackers,

To ping the topic, I'd like to clarify what may be wrong with the idea described here, because I do not see any interest from the community. The topic is related to physical replication. The primary idea is to define the horizon of WAL segments (files) removal based on saved on disk restart LSN values. Now, the WAL segment removal horizon is calculated based on the current restart LSN values of slots, that can not be saved on disk at the time of the horizon calculation. The case take place when a slot is advancing during checkpoint as described earlier in the topic.

Such behaviour is not a problem when slots are used only for physical replication in a conventional way. But it may be a problem when physical slot is used for some other goals. For example, I have an extension which keeps the WAL using physical replication slots. It creates a new physical slot and advances it as needed. After restart, it can use restart lsn of the slot to read WAL from this LSN. In this case, there is no guarantee that restart lsn will point to an existing WAL segment.

The advantage of the current behaviour is that it requires a little bit less WAL to keep. The disadvantage is that physical slots do not guarantee WAL keeping starting from its' restart lsns in general.

I would be happy to get some advice, whether I am on the right or wrong way.  Thank you in advance.

With best regards,

Vitaly

On Thursday, November 07, 2024 16:30 MSK, "Vitaly Davydov" <v.davydov@postgrespro.ru> wrote:
 

Dear Hackers,

 

I'd like to introduce an improved version of my patch (see the attached file). My original idea was to take into account saved on disk restart_lsn (slot→restart_lsn_flushed) for persistent slots when removing WAL segment files. It helps tackle errors like: ERROR: requested WAL segment 000...0AA has already been removed.

 

Improvements:

• flushed_restart_lsn is used only for RS_PERSISTENT slots.
• Save physical slot on disk when advancing only once - if restart_lsn_flushed is invalid. It is needed because slots with invalid restart LSN are not used when calculating oldest LSN for WAL truncation. Once restart_lsn becomes valid, it should be saved to disk immediately to update restart_lsn_flushed.

Regression tests seems to be ok except:

• recovery/t/001_stream_rep.pl (checkpoint is needed)
• recovery/t/019_replslot_limit.pl (it seems, slot was invalidated, some adjustments are needed)
• pg_basebackup/t/020_pg_receivewal.pl (not sure about it)

 

There are some problems:

• More WAL segments may be kept. It may lead to invalidations of slots in some tests (recovery/t/019_replslot_limit.pl). A couple of tests should be adjusted.

 

With best regards,

Vitaly Davydov

On Thursday, October 31, 2024 13:32 MSK, "Vitaly Davydov" <v.davydov@postgrespro.ru> wrote:

 

Sorry, attached the missed patch.

On Thursday, October 31, 2024 13:18 MSK, "Vitaly Davydov" <v.davydov@postgrespro.ru> wrote:

Dear Hackers,

 

I'd like to discuss a problem with replication slots's restart LSN. Physical slots are saved to disk at the beginning of checkpoint. At the end of checkpoint, old WAL segments are recycled or removed from disk, if they are not kept by slot's restart_lsn values.

 

If an existing physical slot is advanced in the middle of checkpoint execution, WAL segments, which are related to saved on disk restart LSN may be removed. It is because the calculation of the replication slot miminal LSN is occured at the end of checkpoint, prior to old WAL segments removal. If to hard stop (pg_stl -m immediate) the postgres instance right after checkpoint and to restart it, the slot's restart_lsn may point to the removed WAL segment. I believe, such behaviour is not good.

 

The doc [0] describes that restart_lsn may be set to the some past value after reload. There is a discussion [1] on pghackers where such behaviour is discussed. The main reason of not flushing physical slots on advancing is a performance reason. I'm ok with such behaviour, except of that the corresponding WAL segments should not be removed.

 

I propose to keep WAL segments by saved on disk (flushed) restart_lsn of slots. Add a new field restart_lsn_flushed into ReplicationSlot structure. Copy restart_lsn to restart_lsn_flushed in SaveSlotToPath. It doesn't change the format of storing the slot contents on disk. I attached a patch. It is not yet complete, but demonstate a way to solve the problem.

 

I reproduced the problem by the following way:

• Add some delay in CheckPointBuffers (pg_usleep) to emulate long checkpoint execution.
• Execute checkpoint and pg_replication_slot_advance right after starting of the checkpoint from another connection.
• Hard restart the server right after checkpoint completion.
• After restart slot's restart_lsn may point to removed WAL segment.

The proposed patch fixes it.

 

[0] https://www.postgresql.org/docs/current/logicaldecoding-explanation.html
[1] https://www.postgresql.org/message-id/flat/059cc53a-8b14-653a-a24d-5f867503b0ee%40postgrespro.ru

 

On 10/31/24 11:18, Vitaly Davydov wrote:
> Dear Hackers,
> 
>  
> 
> I'd like to discuss a problem with replication slots's restart LSN.
> Physical slots are saved to disk at the beginning of checkpoint. At the
> end of checkpoint, old WAL segments are recycled or removed from disk,
> if they are not kept by slot's restart_lsn values.
> 

I agree that if we can lose WAL still needed for a replication slot,
that is a bug. Retaining the WAL is the primary purpose of slots, and we
just fixed a similar issue for logical replication.

> 
> If an existing physical slot is advanced in the middle of checkpoint
> execution, WAL segments, which are related to saved on disk restart LSN
> may be removed. It is because the calculation of the replication slot
> miminal LSN is occured at the end of checkpoint, prior to old WAL
> segments removal. If to hard stop (pg_stl -m immediate) the postgres
> instance right after checkpoint and to restart it, the slot's
> restart_lsn may point to the removed WAL segment. I believe, such
> behaviour is not good.
> 

Not sure I 100% follow, but let me rephrase, just so that we're on the
same page. CreateCheckPoint() does this:

  ... something ...

  CheckPointGuts(checkPoint.redo, flags);

  ... something ...

  RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
                     checkPoint.ThisTimeLineID);

The slots get synced in CheckPointGuts(), so IIUC you're saying if the
slot gets advanced shortly after the sync, the RemoveOldXlogFiles() may
remove still-needed WAL, because we happen to consider a fresh
restart_lsn when calculating the logSegNo. Is that right?

> 
> The doc [0] describes that restart_lsn may be set to the some past value
> after reload. There is a discussion [1] on pghackers where such
> behaviour is discussed. The main reason of not flushing physical slots
> on advancing is a performance reason. I'm ok with such behaviour, except
> of that the corresponding WAL segments should not be removed.
> 

I don't know which part of [0] you refer to, but I guess you're
referring to this:

    The current position of each slot is persisted only at checkpoint,
    so in the case of a crash the slot may return to an earlier LSN,
    which will then cause recent changes to be sent again when the
    server restarts. Logical decoding clients are responsible for
    avoiding ill effects from handling the same message more than once.

Yes, it's fine if we discard the new in-memory restart_lsn value, and we
do this for performance reasons - flushing the slot on every advance
would be very expensive. I haven't read [1] as it's quite long, but I
guess that's what it says.

But we must not make any "permanent" actions based on the unflushed
value, I think. Like, we should not remove WAL segments, for example.

>  
> 
> I propose to keep WAL segments by saved on disk (flushed) restart_lsn of
> slots. Add a new field restart_lsn_flushed into ReplicationSlot
> structure. Copy restart_lsn to restart_lsn_flushed in SaveSlotToPath. It
> doesn't change the format of storing the slot contents on disk. I
> attached a patch. It is not yet complete, but demonstate a way to solve
> the problem.
> 

That seems like a possible fix this, yes. And maybe it's the right one.

> 
> I reproduced the problem by the following way:
> 
>   * Add some delay in CheckPointBuffers (pg_usleep) to emulate long
>     checkpoint execution.
>   * Execute checkpoint and pg_replication_slot_advance right after
>     starting of the checkpoint from another connection.
>   * Hard restart the server right after checkpoint completion.
>   * After restart slot's restart_lsn may point to removed WAL segment.
> 
> The proposed patch fixes it.
> 

I tried to reproduce the issue using a stress test (checkpoint+restart
in a loop), but so far without success :-(

Can you clarify where exactly you added the pg_usleep(), and how long
are the waits you added? I wonder if the sleep is really needed,
considering the checkpoints are spread anyway. Also, what you mean by
"hard reset"?

What confuses me a bit is that we update the restart_lsn (and call
ReplicationSlotsComputeRequiredLSN() to recalculate the global value)
all the time. Walsender does that in PhysicalConfirmReceivedLocation for
example. So we actually see the required LSN to move during checkpoint
very often. So how come we don't see the issues much more often? Surely
I miss something important.

Another option might be that pg_replication_slot_advance() doesn't do
something it should be doing. For example, shouldn't be marking the slot
as dirty?


regards

-- 
Tomas Vondra

On 11/20/24 14:40, Vitaly Davydov wrote:
> Dear Hackers,
> 
>  
> 
> To ping the topic, I'd like to clarify what may be wrong with the idea
> described here, because I do not see any interest from the community.
> The topic is related to physical replication. The primary idea is to
> define the horizon of WAL segments (files) removal based on saved on
> disk restart LSN values. Now, the WAL segment removal horizon is
> calculated based on the current restart LSN values of slots, that can
> not be saved on disk at the time of the horizon calculation. The case
> take place when a slot is advancing during checkpoint as described
> earlier in the topic.
> 

Yeah, a simple way to fix this might be to make sure we don't use the
required LSN value set after CheckPointReplicationSlots() to remove WAL.
AFAICS the problem is KeepLogSeg() gets the new LSN value by:

   keep = XLogGetReplicationSlotMinimumLSN();

Let's say we get the LSN before calling CheckPointGuts(), and then pass
it to KeepLogSeg, so that it doesn't need to get the fresh value.

Wouldn't that fix the issue?

> 
> Such behaviour is not a problem when slots are used only for physical
> replication in a conventional way. But it may be a problem when physical
> slot is used for some other goals. For example, I have an extension
> which keeps the WAL using physical replication slots. It creates a new
> physical slot and advances it as needed. After restart, it can use
> restart lsn of the slot to read WAL from this LSN. In this case, there
> is no guarantee that restart lsn will point to an existing WAL segment.
> 

Yeah.

> 
> The advantage of the current behaviour is that it requires a little bit
> less WAL to keep. The disadvantage is that physical slots do not
> guarantee WAL keeping starting from its' restart lsns in general.
> 

If it's wrong, it doesn't really matter it has some advantages.


regards

-- 
Tomas Vondra

On 11/20/24 18:24, Tomas Vondra wrote:
>
> ...
>
> What confuses me a bit is that we update the restart_lsn (and call
> ReplicationSlotsComputeRequiredLSN() to recalculate the global value)
> all the time. Walsender does that in PhysicalConfirmReceivedLocation for
> example. So we actually see the required LSN to move during checkpoint
> very often. So how come we don't see the issues much more often? Surely
> I miss something important.
> 

This question "How come we don't see this more often?" kept bugging me,
and the answer is actually pretty simple.

The restart_lsn can move backwards after a hard restart (for the reasons
explained), but physical replication does not actually rely on that. The
replica keeps track of the LSN it received (well, it uses the same LSN),
and on reconnect it sends the startpoint to the primary. And the primary
just proceeds use that instead of the (stale) restart LSN for the slot.
And the startpoint is guaranteed (I think) to be at least restart_lsn.

AFAICS this would work for pg_replication_slot_advance() too, that is if
you remember the last LSN the slot advanced to, it should be possible to
advance to it just fine. Of course, it requires a way to remember that
LSN, which for a replica is not an issue. But this just highlights we
can't rely on restart_lsn for this purpose.


(Apologies if this repeats something obvious, or something you already
said, Vitaly.)

regards

-- 
Tomas Vondra

On 11/20/24 23:19, Tomas Vondra wrote:
> On 11/20/24 18:24, Tomas Vondra wrote:
>>
>> ...
>>
>> What confuses me a bit is that we update the restart_lsn (and call
>> ReplicationSlotsComputeRequiredLSN() to recalculate the global value)
>> all the time. Walsender does that in PhysicalConfirmReceivedLocation for
>> example. So we actually see the required LSN to move during checkpoint
>> very often. So how come we don't see the issues much more often? Surely
>> I miss something important.
>>
> 
> This question "How come we don't see this more often?" kept bugging me,
> and the answer is actually pretty simple.
> 
> The restart_lsn can move backwards after a hard restart (for the reasons
> explained), but physical replication does not actually rely on that. The
> replica keeps track of the LSN it received (well, it uses the same LSN),
> and on reconnect it sends the startpoint to the primary. And the primary
> just proceeds use that instead of the (stale) restart LSN for the slot.
> And the startpoint is guaranteed (I think) to be at least restart_lsn.
> 
> AFAICS this would work for pg_replication_slot_advance() too, that is if
> you remember the last LSN the slot advanced to, it should be possible to
> advance to it just fine. Of course, it requires a way to remember that
> LSN, which for a replica is not an issue. But this just highlights we
> can't rely on restart_lsn for this purpose.
> 

I kept thinking about this (sorry it's this incremental), particularly
if this applies to logical replication too. And AFAICS it does not, or
at least not to this extent.

For streaming, the subscriber sends the startpoint (just like physical
replication), so it should be protected too.

But then there's the SQL API - pg_logical_slot_get_changes(). And it
turns out it ends up syncing the slot to disk pretty often, because for
RUNNING_XACTS we call LogicalDecodingProcessRecord() + standby_decode(),
which ends up calling SaveSlotToDisk(). And at the end we call
LogicalConfirmReceivedLocation() for good measure, which saves the slot
too, just to be sure.

FWIW I suspect this still is not perfectly safe, because we may still
crash / restart before the updated data.restart_lsn makes it to disk,
but after it was already used to remove old WAL, although that's
probably harder to hit. With streaming the subscriber will still send us
the new startpoint, so that should not fail I think. But with the SQL
API we probably can get into the "segment already removed" issues.

I haven't tried reproducing this yet, I guess it should be possible
using the injection points. Not sure when I get to this, though.


In any case, doesn't this suggest SaveSlotToDisk() really is not that
expensive, if we do it pretty often for logical replication? Which was
presented as the main reason why pg_replication_slot_advance() doesn't
do that. Maybe it should?

If the advance is substantial I don't think it really matters, because
there simply can't be that many of large advances. It amortizes, in a
way. But even with smaller advances it should be fine, I think - if the
goal is to not remove WAL prematurely, it's enough to flush when we move
to the next segment.


regards

-- 
Tomas Vondra

Hi Tomas,

Thank you for the reply and your interest to the investigation.

On Wednesday, November 20, 2024 20:24 MSK, Tomas Vondra <tomas@vondra.me> wrote:
 

> If an existing physical slot is advanced in the middle of checkpoint
> execution, WAL segments, which are related to saved on disk restart LSN
> may be removed. It is because the calculation of the replication slot
> miminal LSN is occured at the end of checkpoint, prior to old WAL
> segments removal. If to hard stop (pg_stl -m immediate) the postgres
> instance right after checkpoint and to restart it, the slot's
> restart_lsn may point to the removed WAL segment. I believe, such
> behaviour is not good.
> 

Not sure I 100% follow, but let me rephrase, just so that we're on the
same page. CreateCheckPoint() does this:

... something ...

CheckPointGuts(checkPoint.redo, flags);

... something ...

RemoveOldXlogFiles(_logSegNo, RedoRecPtr, recptr,
checkPoint.ThisTimeLineID);

The slots get synced in CheckPointGuts(), so IIUC you're saying if the
slot gets advanced shortly after the sync, the RemoveOldXlogFiles() may
remove still-needed WAL, because we happen to consider a fresh
restart_lsn when calculating the logSegNo. Is that right?

They key action here is to restart the instance with -m immediate (or kill it and start it again) right after checkpoint. After restart, the slot's restart_lsn will be read from the disk and address to the removed WAL segment, if it's LSN was davanced enought to switch to a new WAL segment.

I tried to reproduce the issue using a stress test (checkpoint+restart
in a loop), but so far without success :-(

Can you clarify where exactly you added the pg_usleep(), and how long
are the waits you added? I wonder if the sleep is really needed,
considering the checkpoints are spread anyway. Also, what you mean by
"hard reset"?
 

I added pg_usleep as show below (in CheckPointBuffers function):

CheckPointBuffers(int flags)
{
    BufferSync(flags);
+    pg_usleep(10000000);
}
 

Below is the instruction how I run my test (pg_usleep should be added to the code):

CONSOLE> initdb -D pgdata
CONSOLE> pg_ctl -D pgdata -l logfile start
... open two psql terminals and connect to the database (lets call them as PSQL-1, PSQL-2)
PSQL-1>  select pg_create_physical_replication_slot('myslot', true, false);
CONSOLE> pgbench -i -s 10 postgres # create some WAL records
PSQL-1>  checkpoint; -- press ENTER key and go to PSQL-2 console and execute next line in 1 second
PSQL-2>  select pg_replication_slot_advance('myslot', pg_current_wal_lsn()); -- advance repslot during checkpoint
... wait for checkpoint to complete
CONSOLE> pg_ctl -D pgdata -m immediate stop 
CONSOLE> pg_ctl -D pgdata start
PSQL-1>  \c
PSQL-1>  create extension pg_walinspect;
PSQL-1>  select pg_get_wal_record_info(restart_lsn) from pg_replication_slots where slot_name = 'myslot';
ERROR:  requested WAL segment pg_wal/000000010000000000000001 has already been removed

I'm trying to create a perl test to reproduce it. Please, give me some time to create the test script.

I kept thinking about this (sorry it's this incremental), particularly
if this applies to logical replication too. And AFAICS it does not, or
at least not to this extent.

 

Yes, it is not applied to logical replication, because logical slot is synced when advancing.

eah, a simple way to fix this might be to make sure we don't use the
required LSN value set after CheckPointReplicationSlots() to remove WAL.
AFAICS the problem is KeepLogSeg() gets the new LSN value by:

keep = XLogGetReplicationSlotMinimumLSN();

Let's say we get the LSN before calling CheckPointGuts(), and then pass
it to KeepLogSeg, so that it doesn't need to get the fresh value.

 

Yes, it is another solution and it can fix the problem. The question - which solution to choose. Well, I prefer to add a new in-memory state variable in the slot structure. Such variable may be useful if we want to check whether the slot data is synced or not. The calculation of the keep value before CheckPointGuts(), IMHO, requires to change signatures of a number of functions. I may prepare a new patch where your solution is implemented.

I'm sorry, if I missed to answer to some other questions. I will answer later.

With best regards,

Vitaly

Dear Hackers,

Let me please introduce a new version of the patch.

Patch description:

The slot data is flushed to the disk at the beginning of checkpoint. If
an existing slot is advanced in the middle of checkpoint execution, its
advanced restart LSN is taken to calculate the oldest LSN for WAL
segments removal at the end of checkpoint. If the node is restarted just
after the checkpoint, the slots data will be read from the disk at
recovery with the oldest restart LSN which can refer to removed WAL
segments.

The patch introduces a new in-memory state for slots -
flushed_restart_lsn which is used to calculate the oldest LSN for WAL
segments removal. This state is updated every time with the current
restart_lsn at the moment, when the slot is saving to disk.

With best regards,
Vitaly

Hi, Vitaly!

On Mon, Mar 3, 2025 at 5:12 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> The slot data is flushed to the disk at the beginning of checkpoint. If
> an existing slot is advanced in the middle of checkpoint execution, its
> advanced restart LSN is taken to calculate the oldest LSN for WAL
> segments removal at the end of checkpoint. If the node is restarted just
> after the checkpoint, the slots data will be read from the disk at
> recovery with the oldest restart LSN which can refer to removed WAL
> segments.
>
> The patch introduces a new in-memory state for slots -
> flushed_restart_lsn which is used to calculate the oldest LSN for WAL
> segments removal. This state is updated every time with the current
> restart_lsn at the moment, when the slot is saving to disk.

Thank you for your work on this subject.   I think generally your
approach is correct.  When we're truncating the WAL log, we need to
reply on the position that would be used in the case of server crush.
That is the position flushed to the disk.

While your patch is generality looks good, I'd like make following notes:

1) As ReplicationSlotsComputeRequiredLSN() is called each time we need
to advance the position of WAL needed by replication slots, the usage
pattern probably could be changed.  Thus, we probably need to call
ReplicationSlotsComputeRequiredLSN() somewhere after change of
restart_lsn_flushed while restart_lsn is not changed.  And probably
can skip ReplicationSlotsComputeRequiredLSN() in some cases when only
restart_lsn is changed.
2) I think it's essential to include into the patch test caches which
fail without patch.  You could start from integrating [1] test into
your patch, and then add more similar tests for different situations.

Links.
1.  https://www.postgresql.org/message-id/e3ac0535-e7a2-4a96-9b36-9f765e9cfec5%40vondra.me

------
Regards,
Alexander Korotkov
Supabase

Hi Alexander,

Thank you for the review. I apologize for a late reply. I missed your email.

> 1) As ReplicationSlotsComputeRequiredLSN() is called each time we need
> to advance the position of WAL needed by replication slots, the usage
> pattern probably could be changed.  Thus, we probably need to call
> ReplicationSlotsComputeRequiredLSN() somewhere after change of
> restart_lsn_flushed while restart_lsn is not changed.  And probably
> can skip ReplicationSlotsComputeRequiredLSN() in some cases when only
> restart_lsn is changed.

Yes, it is a good idea for investigation, thank you! I guess, It may work for
persistent slots but I'm not sure about other types of slots (ephemeral and
temporary). I have no clear understanding of consequences at the moment. I
propose to postpone it for future, because the proposed changes will me more
invasive.

> 2) I think it's essential to include into the patch test caches which
> fail without patch.  You could start from integrating [1] test into
> your patch, and then add more similar tests for different situations.

The problem with TAP tests - it is hard to reproduce without injection points.
The Tomas Vondra's tests create two new injection points. I have to add more
injection points for new tests as well. Injection points help to test the code
but make the code unreadable. I'm not sure, what is the policy of creating
injection points? Personally, I would not like to add new injection points
only to check this particular rare case. I'm trying to create a test without
injection points that should fail occasionally, but I haven't succeeded at
the moment.

I have a question - is there any interest to backport the solution into
existing major releases? I can prepare a patch where restart_lsn_flushed stored
outside of ReplicationSlot structure and doesn't affect the existing API.

With best regards,
Vitaly

On Friday, April 04, 2025 06:22 MSK, Alexander Korotkov <aekorotkov@gmail.com> wrote:

> Hi, Vitaly!
>
> On Mon, Mar 3, 2025 at 5:12 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> > The slot data is flushed to the disk at the beginning of checkpoint. If
> > an existing slot is advanced in the middle of checkpoint execution, its
> > advanced restart LSN is taken to calculate the oldest LSN for WAL
> > segments removal at the end of checkpoint. If the node is restarted just
> > after the checkpoint, the slots data will be read from the disk at
> > recovery with the oldest restart LSN which can refer to removed WAL
> > segments.
> >
> > The patch introduces a new in-memory state for slots -
> > flushed_restart_lsn which is used to calculate the oldest LSN for WAL
> > segments removal. This state is updated every time with the current
> > restart_lsn at the moment, when the slot is saving to disk.
>
> Thank you for your work on this subject.   I think generally your
> approach is correct.  When we're truncating the WAL log, we need to
> reply on the position that would be used in the case of server crush.
> That is the position flushed to the disk.
>
> While your patch is generality looks good, I'd like make following notes:
>
> 1) As ReplicationSlotsComputeRequiredLSN() is called each time we need
> to advance the position of WAL needed by replication slots, the usage
> pattern probably could be changed.  Thus, we probably need to call
> ReplicationSlotsComputeRequiredLSN() somewhere after change of
> restart_lsn_flushed while restart_lsn is not changed.  And probably
> can skip ReplicationSlotsComputeRequiredLSN() in some cases when only
> restart_lsn is changed.
> 2) I think it's essential to include into the patch test caches which
> fail without patch.  You could start from integrating [1] test into
> your patch, and then add more similar tests for different situations.
>
> Links.
> 1.  https://www.postgresql.org/message-id/e3ac0535-e7a2-4a96-9b36-9f765e9cfec5%40vondra.me
>
> ------
> Regards,
> Alexander Korotkov
> Supabase

On Thu, Apr 24, 2025 at 5:32 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> Thank you for the review. I apologize for a late reply. I missed your email.
>
> > 1) As ReplicationSlotsComputeRequiredLSN() is called each time we need
> > to advance the position of WAL needed by replication slots, the usage
> > pattern probably could be changed.  Thus, we probably need to call
> > ReplicationSlotsComputeRequiredLSN() somewhere after change of
> > restart_lsn_flushed while restart_lsn is not changed.  And probably
> > can skip ReplicationSlotsComputeRequiredLSN() in some cases when only
> > restart_lsn is changed.
>
> Yes, it is a good idea for investigation, thank you! I guess, It may work for
> persistent slots but I'm not sure about other types of slots (ephemeral and
> temporary). I have no clear understanding of consequences at the moment. I
> propose to postpone it for future, because the proposed changes will me more
> invasive.

Yes, that's different for different types of slots.  So, removing
ReplicationSlotsComputeRequiredLSN() doesn't look safe.  But at least,
we need to analyze if we need to add extra calls.

> > 2) I think it's essential to include into the patch test caches which
> > fail without patch.  You could start from integrating [1] test into
> > your patch, and then add more similar tests for different situations.
>
> The problem with TAP tests - it is hard to reproduce without injection points.
> The Tomas Vondra's tests create two new injection points. I have to add more
> injection points for new tests as well. Injection points help to test the code
> but make the code unreadable. I'm not sure, what is the policy of creating
> injection points? Personally, I would not like to add new injection points
> only to check this particular rare case. I'm trying to create a test without
> injection points that should fail occasionally, but I haven't succeeded at
> the moment.

I don't know if there is an explicit policy.  I think we just add them
as needed to reproduce important situations in TAP tests.  So, feel
free to them as many as you want to reproduce all the problematic
situations.  During review we can find if they seem too many, but
don't bother about this at present stage.

> I have a question - is there any interest to backport the solution into
> existing major releases?

As long as this is the bug, it should be backpatched to all supported
affected releases.

> I can prepare a patch where restart_lsn_flushed stored
> outside of ReplicationSlot structure and doesn't affect the existing API.

Yes, please!

------
Regards,
Alexander Korotkov
Supabase

On Mon, Apr 28, 2025 at 8:17 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> > I have a question - is there any interest to backport the solution into
> > existing major releases?
>
> As long as this is the bug, it should be backpatched to all supported
> affected releases.

Yes, but I think we cannot back-patch the proposed fix to back
branches as it changes the ReplicationSlot struct defined in slot.h,
which breaks ABI compatibility.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com

On Tue, Apr 29, 2025 at 4:03 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>
> On Mon, Apr 28, 2025 at 8:17 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> >
> > > I have a question - is there any interest to backport the solution into
> > > existing major releases?
> >
> > As long as this is the bug, it should be backpatched to all supported
> > affected releases.
>
> Yes, but I think we cannot back-patch the proposed fix to back
> branches as it changes the ReplicationSlot struct defined in slot.h,
> which breaks ABI compatibility.

Yes, and I think Vitaly already proposed to address this issue.  This
aspect also needs to be carefully reviewed for sure.

On Thu, Apr 24, 2025 at 5:32 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> I can prepare a patch where restart_lsn_flushed stored
> outside of ReplicationSlot structure and doesn't affect the existing API.

------
Regards,
Alexander Korotkov
Supabase

On Mon, Apr 28, 2025 at 6:39 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> On Tue, Apr 29, 2025 at 4:03 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> >
> > On Mon, Apr 28, 2025 at 8:17 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> > >
> > > > I have a question - is there any interest to backport the solution into
> > > > existing major releases?
> > >
> > > As long as this is the bug, it should be backpatched to all supported
> > > affected releases.
> >
> > Yes, but I think we cannot back-patch the proposed fix to back
> > branches as it changes the ReplicationSlot struct defined in slot.h,
> > which breaks ABI compatibility.
>
> Yes, and I think Vitaly already proposed to address this issue.  This
> aspect also needs to be carefully reviewed for sure.
>
> On Thu, Apr 24, 2025 at 5:32 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> > I can prepare a patch where restart_lsn_flushed stored
> > outside of ReplicationSlot structure and doesn't affect the existing API.

Oh, I totally missed this part. Sorry for making noise. I'll review
the patch once submitted.

Regarding the proposed patch, I think we can somewhat follow
last_saved_confirmed_flush field of ReplicationSlot. For example, we
can set restart_lsn_flushed when restoring the slot from the disk,
etc.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com

Dear All,

Thank you for the attention to the patch. I updated a patch with a better
solution for the master branch which can be easily backported to the other
branches as we agree on the final solution.

Two tests are introduced which are based on Tomas Vondra's test for logical slots
with injection points from the discussion (patches: [1], [2], [3]). Tests
are implemented as module tests in src/test/modules/test_replslot_required_lsn
directory. I slightly modified the original test for logical slots and created a
new simpler test for physical slots without any additional injection points.

I prepared a new solution (patch [4]) which is also based on Tomas Vondra's
proposal. With a fresh eye, I realized that it can fix the issue as well. It is
easy and less invasive to implement. The new solution differs from my original
solution: it is backward compatible (doesn't require any changes in ReplicationSlot
structure). My original solution can be backward compatible as well if to
allocate flushed_restart_lsn in a separate array in shmem, not in the
ReplicationSlot structure, but I believe the new solution is the better one. If
you still think that my previous solution is the better (I don't think so), I
will prepare a backward compatible patch with my previous solution.

I also proposed one more commit (patch [5]) which removes unnecessary calls of
ReplicationSlotsComputeRequiredLSN function which seems to be redundant. This
function updates the oldest required LSN for slots and it is called every time
when slots' restart_lsn is changed. Once, we use the oldest required LSN in
CreateCheckPoint/CreateRestartPoint to remove old WAL segments, I believe, there
is no need to calculate the oldest value immediately when the slot is advancing
and in other cases when restart_lsn is changed. It may affect on
GetWALAvailability function because the oldest required LSN will be not up to
date, but this function seems to be used in the system view
pg_get_replication_slots and doesn't affect the logic of old WAL segments
removal. I also have some doubts concerning advancing of logical replication
slots: the call of ReplicationSlotsComputeRequiredLSN was removed. Not sure, how
it can affect on ReplicationSlotsComputeRequiredXmin. This commit is not
necessary for the fix but I think it is worth to consider. It may be dropped or
applied only to the master branch.

This patch can be easily backported to the major release branches. I will
quickly prepare the patches for the major releases as we agree on the final
solution.

I apologize for such too late change in patch when it is already on commitfest.
I'm not well experienced yet in the internals of PostgreSQL at the moment,
sometimes the better solution needs some time to grow. In doing we learn :)

[1] 0001-Add-injection-points-to-test-replication-slot-advanc.v2.patch
[2] 0002-Add-TAP-test-to-check-logical-repl-slot-advance-duri.v2.patch
[3] 0003-Add-TAP-test-to-check-physical-repl-slot-advance-dur.v2.patch
[4] 0004-Keep-WAL-segments-by-slot-s-flushed-restart-LSN.v2.patch
[5] 0005-Remove-redundant-ReplicationSlotsComputeRequiredLSN-.v2.patch

With best regards,
Vitaly

Hi Alexander,

Thank you very much for the review!

> The patchset doesn't seem to build after 371f2db8b0, which adjusted
> the signature of the INJECTION_POINT() macro.  Could you, please,
> update the patchset accordingly.

I've updated the patch (see attached). Thanks.

> I see in 0004 patch we're calling XLogGetReplicationSlotMinimumLSN()
> before slots synchronization then use it for WAL truncation.
> Generally looks good.  But what about the "if
> (InvalidateObsoleteReplicationSlots(...))" branch?  It calls
> XLogGetReplicationSlotMinimumLSN() again.  Why would the value
> obtained from the latter call reflect slots as they are synchronized
> to the disk?

In patch 0004 I call XLogGetReplicationSlotMinimumLSN() again to keep the old
behaviour - this function was called in KeepLogSeg prior to my change. I also
call CheckPointReplicationSlots at the next line to save the invalidated and
other dirty slots on disk again to make sure, the new oldest LSN is in sync.

The problem I tried to solve in this if-branch is to fix test
src/test/recovery/t/019_replslot_limit.pl which was failed because the WAL was
not truncated enought for the test to pass ok. In general, this branch is not
necessary and we may fix the test by calling checkpoint twice (please, see the
alternative.rej patch for this case). If you think, we should incorporate this
new change, I'm ok to do it. But the WAL will be truncating more lazily.

Furthermore, I think we can save slots on disk right after invalidation, not in
CheckPointGuts to avoid saving invalidated slots twice.

With best regards,
Vitaly

Hi, Amit!

Thank you for your attention to this patchset!

On Sat, May 24, 2025 at 2:15 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> On Sat, May 24, 2025 at 4:08 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> >
> > I spend more time on this.  The next revision is attached.  It
> > contains revised comments and other cosmetic changes.  I'm going to
> > backpatch 0001 to all supported branches,
> >
>
> Is my understanding correct that we need 0001 because
> PhysicalConfirmReceivedLocation() doesn't save the slot to disk after
> changing the slot's restart_lsn?

Yes.  Also, even if it would save slot to the disk, there is still
race condition that concurrent checkpoint could use updated value from
the shared memory to clean old WAL segments, and then crash happens
before we managed to write the slot to the disk.

> If so, shouldn't the comments (One
> could argue that the slot should be saved to disk now, but that'd be
> energy wasted - the worst thing lost information could cause here is
> to give wrong information in a statistics view) in
> PhysicalConfirmReceivedLocation() be changed to mention the risk of
> not saving the slot?
>
> Also, after 0001, even the same solution will be true for logical
> slots as well, where we are already careful to save the slot to disk
> if its restart_lsn is changed, see LogicalConfirmReceivedLocation().
> So, won't that effort be wasted? Even if we don't want to do anything
> about it (which doesn't sound like a good idea), we should note that
> in comments somewhere.

I have added the comments about both points in the attached revision
of the patchset.

------
Regards,
Alexander Korotkov
Supabase

On Sat, May 24, 2025 at 6:59 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> Hi, Amit!
>
> Thank you for your attention to this patchset!
>
> On Sat, May 24, 2025 at 2:15 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > On Sat, May 24, 2025 at 4:08 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> > >
> > > I spend more time on this.  The next revision is attached.  It
> > > contains revised comments and other cosmetic changes.  I'm going to
> > > backpatch 0001 to all supported branches,
> > >
> >
> > Is my understanding correct that we need 0001 because
> > PhysicalConfirmReceivedLocation() doesn't save the slot to disk after
> > changing the slot's restart_lsn?
>
> Yes.  Also, even if it would save slot to the disk, there is still
> race condition that concurrent checkpoint could use updated value from
> the shared memory to clean old WAL segments, and then crash happens
> before we managed to write the slot to the disk.
>

How can that happen, if we first write the updated value to disk and
then update the shared memory as we do in
LogicalConfirmReceivedLocation?

BTW, won't there be a similar problem with physical slot's xmin
computation as well? In PhysicalReplicationSlotNewXmin(), after
updating the slot's xmin computation, we mark it as dirty and update
shared memory values. Now, say after checkpointer writes these xmin
values to disk, walsender receives another feedback message and
updates the slot's xmin values. Now using these updated shared memory
values, vacuum removes the rows, however, a restart will show the
older xmin values in the slot, which mean vacuum would have removed
the required rows before restart.

As per my understanding, neither the xmin nor the LSN problem exists
for logical slots. I am pointing this out to indicate we may need to
think of a different solution for physical slots, if these are
problems only for physical slots.

--
With Regards,
Amit Kapila.

Dear Alexander, Amit, All

> Amit wrote:
> > Is my understanding correct that we need 0001 because
> > PhysicalConfirmReceivedLocation() doesn't save the slot to disk after
> > changing the slot's restart_lsn?
>
> Yes.  Also, even if it would save slot to the disk, there is still
> race condition that concurrent checkpoint could use updated value from
> the shared memory to clean old WAL segments, and then crash happens
> before we managed to write the slot to the disk.
>
> How can that happen, if we first write the updated value to disk and
> then update the shared memory as we do in
> LogicalConfirmReceivedLocation?

I guess, that the problem with logical slots still exist. Please, see the tap
test: src/test/recovery/t/046_logical_slot.pl from the v6 version of the patch.
A race condition may happen when logical slot's restart_lsn was changed but not
yet written to the disk. Imagine, there is another physical slot which is
advanced at this moment. It recomputes oldest min LSN and takes into account
changed but not saved to disk restart_lsn of the logical slot. We come to the
situation when the WAL segment for the logical slot's restart_lsn may be
truncated after immediate restart.

I'm not sure what may happen with two checkpoints which execute in parallel, but
I would say that the patch 0001 guarantees that every checkpoint run will trim
the WAL segments based on the already saved on disk restart LSNs of slots. The
rule to trim the WAL by saved slot's restart_lsn will not be violated.

> Amit wrote:
> As per my understanding, neither the xmin nor the LSN problem exists
> for logical slots. I am pointing this out to indicate we may need to
> think of a different solution for physical slots, if these are
> problems only for physical slots.

As I've already told, it indirectly affects the logical slots as well.

> Alexander wrote:
> I spend more time on this.  The next revision is attached.  It
> contains revised comments and other cosmetic changes.  I'm going to
> backpatch 0001 to all supported branches, and 0002 to 17 where
> injection points were introduced.

Alexander, thank you for polishing the patch. Just my opinion, I would prefer
to put tests before the fix due to reason that you can reproduce the problem
when simply checkout the commit with tests. Once, the tests are after the fix
you are not able to do this way. Anyway, I'm ok with your changes. Thank you!

I did some changes in the patch (v7 is attached):
* Removed modules/test_replslot_required_lsn directory. It is not needed anymore,
once you've moved test files to another directory.
* Renamed tests to 046_checkpoint_logical_slot.pl, 047_checkpoint_physical_slot.pl.
I believe, such names are more descriptive.

Please, consider these changes.

With best regards,
Vitaly

On Mon, May 26, 2025 at 3:52 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
>
> Dear Alexander, Amit, All
>
> > Amit wrote:
> > > Is my understanding correct that we need 0001 because
> > > PhysicalConfirmReceivedLocation() doesn't save the slot to disk after
> > > changing the slot's restart_lsn?
> >
> > Yes.  Also, even if it would save slot to the disk, there is still
> > race condition that concurrent checkpoint could use updated value from
> > the shared memory to clean old WAL segments, and then crash happens
> > before we managed to write the slot to the disk.
> >
> > How can that happen, if we first write the updated value to disk and
> > then update the shared memory as we do in
> > LogicalConfirmReceivedLocation?
>
> I guess, that the problem with logical slots still exist. Please, see the tap
> test: src/test/recovery/t/046_logical_slot.pl from the v6 version of the patch.
> A race condition may happen when logical slot's restart_lsn was changed but not
> yet written to the disk. Imagine, there is another physical slot which is
> advanced at this moment. It recomputes oldest min LSN and takes into account
> changed but not saved to disk restart_lsn of the logical slot. We come to the
> situation when the WAL segment for the logical slot's restart_lsn may be
> truncated after immediate restart.
>

Okay, so I was missing the point that the physical slots can consider
the updated value of the logical slot's restart_lsn. The point I was
advocating for logical slots sanctity was when no physical slots are
involved. When updating replicationSlotMinLSN value in shared memory,
the logical slot machinery took care that the value we use should be
flushed to disk. One can argue that we should improve physical slots
machinery so that it also takes care to write the slot to disk before
updating the replicationSlotMinLSN, which is used to remove WAL. I
understand that the downside is physical slots will be written to disk
with a greater frequency, which will not be good from the performance
point of view, but can we think of doing it for the period when a
checkpoint is in progress? OTOH, if we don't want to adjust physical
slot machinery, it seems saving the logical slots to disk immediately
when its restart_lsn is updated is a waste of effort after your patch,
no? If so, why are we okay with that?

I understand that your proposed patch fixes the reported problem but I
am slightly afraid that the proposed solution is not a good idea w.r.t
logical slots so I am trying to see if there are any other alternative
ideas to fix this issue.

--
With Regards,
Amit Kapila.

Dear Amit,

> OTOH, if we don't want to adjust physical
> slot machinery, it seems saving the logical slots to disk immediately
> when its restart_lsn is updated is a waste of effort after your patch,
> no? If so, why are we okay with that?

I agree, that saving logical slots at advance is a possible waste of effort. But
I don't understand original ideas behind it. I haven't touched it to make
the minimal patch which should not break the existing functionality.

We trim WAL in checkpoint (or restart point) operation only. The slots'
restart_lsn is used to keep the wal from truncation. I believe, we need to
compute the slots' oldest lsn as the minimal value of restart_lsn values only
when executing checkpoint (or restart point). I guess, it doesn't depend on
slot's type (logical or physical). We have 0003 patch to fix it.

I haven't deeply investigated yet slot's xmin values but I guess the xmin values
are a different story than restart_lsn. It is used to avoid tuple deletions by
vacuum and it is updated by a different way. I can't say that
LogicalConfirmReceivedLocation is the right place to update saved on disk xmin
values. I would propose to update these values in SaveSlotToPath under some lock
to avoid concurrent reads of unsaved values or do in a checkpoint like as for
restart_lsn. We may investigate and improve it in an another patch.

With best regards,
Vitaly

On Mon, May 26, 2025 at 2:43 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
>
> On Mon, May 26, 2025 at 3:52 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
> >
> > Dear Alexander, Amit, All
> >
> > > Amit wrote:
> > > > Is my understanding correct that we need 0001 because
> > > > PhysicalConfirmReceivedLocation() doesn't save the slot to disk after
> > > > changing the slot's restart_lsn?
> > >
> > > Yes.  Also, even if it would save slot to the disk, there is still
> > > race condition that concurrent checkpoint could use updated value from
> > > the shared memory to clean old WAL segments, and then crash happens
> > > before we managed to write the slot to the disk.
> > >
> > > How can that happen, if we first write the updated value to disk and
> > > then update the shared memory as we do in
> > > LogicalConfirmReceivedLocation?
> >
> > I guess, that the problem with logical slots still exist. Please, see the tap
> > test: src/test/recovery/t/046_logical_slot.pl from the v6 version of the patch.
> > A race condition may happen when logical slot's restart_lsn was changed but not
> > yet written to the disk. Imagine, there is another physical slot which is
> > advanced at this moment. It recomputes oldest min LSN and takes into account
> > changed but not saved to disk restart_lsn of the logical slot. We come to the
> > situation when the WAL segment for the logical slot's restart_lsn may be
> > truncated after immediate restart.
> >
>
> Okay, so I was missing the point that the physical slots can consider
> the updated value of the logical slot's restart_lsn. The point I was
> advocating for logical slots sanctity was when no physical slots are
> involved. When updating replicationSlotMinLSN value in shared memory,
> the logical slot machinery took care that the value we use should be
> flushed to disk. One can argue that we should improve physical slots
> machinery so that it also takes care to write the slot to disk before
> updating the replicationSlotMinLSN, which is used to remove WAL. I
> understand that the downside is physical slots will be written to disk
> with a greater frequency, which will not be good from the performance
> point of view, but can we think of doing it for the period when a
> checkpoint is in progress?

That could cause replication slowdown while checkpointing is
in-progress.  This is certainly better than slowing down the
replication permanently, but still doesn't look good.

> OTOH, if we don't want to adjust physical
> slot machinery, it seems saving the logical slots to disk immediately
> when its restart_lsn is updated is a waste of effort after your patch,
> no? If so, why are we okay with that?

I don't think so.  I think the reason why logical slots are synced to
disk immediately after update is that logical changes are not
idempotent (you can't safely apply the same change twice) unlike
physical block-level changes.  This is why logical slots need to be
synced to prevent double replication of same changes, which could
lead, for example, to double insertion.

> I understand that your proposed patch fixes the reported problem but I
> am slightly afraid that the proposed solution is not a good idea w.r.t
> logical slots so I am trying to see if there are any other alternative
> ideas to fix this issue.

I don't understand exact concerns about this fix.  For sure, we can
try to implement a fix hacking LogicalConfirmReceivedLocation() and
PhysicalConfirmReceivedLocation().  But that would be way more
cumbersome, especially if we have to keep ABI compatibility.  Also, it
doesn't seem to me that either LogicalConfirmReceivedLocation() or
PhysicalConfirmReceivedLocation() currently try to address this issue:
LogicalConfirmReceivedLocation() implements immediate sync for
different reasons.

------
Regards,
Alexander Korotkov
Supabase

On Thu, May 29, 2025 at 5:29 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> On Tue, May 27, 2025 at 2:26 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > Yeah, we should be able to change ABI during beta, but I can't comment
> > on the idea of effective_restart_lsn without seeing the patch or a
> > detailed explanation of this idea.
>
> Could you, please, check the patch [1].  It implements this idea
> except it names new field restart_lsn_flushed instead of
> effective_restart_lsn.
>

This appears to be a better direction than the other patch, at least
for HEAD. I noticed a few points while looking at the patch.

1. restart_lsn_flushed: Can we name it as last_saved_restart_lsn based
on existing variable last_saved_confirmed_flush?
2. There are no comments as to why this is considered only for
persistent slots when CheckPointReplicationSlots doesn't have any such
check.
3. Please see if it makes sense to copy it in the copy_replication_slot.

Apart from these, I am not sure if there are still any pending
comments in the thread to be handled for this patch, so please see to
avoid missing anything.

> > Now, you see my point related to restart_lsn computation for logical
> > slots, it is better to also do some analysis of the problem related to
> > xmin I have highlighted in one of my previous emails [1]. I see your
> > response to it, but I feel someone needs to give it a try by writing a
> > test and see the behavior. I am saying because logical slots took
> > precaution of flushing to disk before updating shared values of xmin
> > for a reason, whereas similar precautions are not taken for physical
> > slots, so there could be a problem with that computation as well.
>
> I see LogicalConfirmReceivedLocation() performs correctly while
> updating effective_catalog_xmin only after syncing the slot to the
> disk.  I don't see how effective_xmin gets updates with the logical
> replication progress though.  Could you get me some clue on this,
> please?
>

As per my understanding, for logical slots, effective_xmin is only set
during the initial copy phase (or say if one has to export a
snapshot), after that, its value won't change. Please read the
comments in CreateInitDecodingContext() where we set its value. If you
still have questions about it, we can discuss further.

--
With Regards,
Amit Kapila.

Dear Alexander, Amit

Alexander Korotkov wrote:
> Also, I've changed ReplicationSlotsComputeRequiredLSN() call to
> CheckPointReplicationSlots() to update required LSN after
> SaveSlotToPath() updated last_saved_restart_lsn.  This helps to pass
> checks in 001_stream_rep.pl without additional hacks.

Thank you for the improvement and patch preparation. I confirm the test is
passed without additional hacks now.

I sill do not understand why this solution is favored. It is, in my opinion,
a non backward-compatible solution. In any case, I'm ok to go with this patch.
If needed, I may prepare a backward-compatible solution where
last_saved_restart_lsn values will be in an another place of the shmem, rather
than in ReplicationSlot struct.

I still would like to add my 5 cents to the discussion.

The purpose of the xmin value is to prevent tuples from vacuuming. Slots'
restart_lsn values are used to calculate the oldest lsn to keep WAL segments
from removal in checkpoint. These processes are pretty independent.

The logical slots are advanced in 2 steps. At the first step, the logical
decoding stuff periodically sets consistent candidate values for catalog_xmin and
restart_lsn. At the second step, when LogicalConfirmReceivedLocation is called,
the candidate values are assigned on catalog_xmin and restart_lsn values based
on the confirmed lsn value. The slot is saved with these consistent values.
It is important, that the candidate values are consistent, decoding guarantees
it. In case of a crash, we should guarantee that the loaded from the disk
catalog_xmin and restart_lsn values are consistent and valid for logical slots.
LogicalConfirmReceivedLocation function keeps this consistency by updating them
from consistent candidate values in a single operation.

We have to guarantee that we use saved to disk values to calculate xmin horizon
and slots' oldest lsn. For this purpose, effective_catalog_xmin is used. We
update effective_catalog_xmin in LogicalConfirmReceivedLocation just
after saving slot to disk. Another place where we update effective_catalog_xmin
is when walsender receives hot standby feedback message.

Once, we have two independent processes (vacuuming, checkpoint), we can calculate
xmin horizon and oldest WAL lsn values independently (at different times) from
the saved to disk values. Note, these values are updated in a non atomic way.

The xmin value is set when the node receives hot standby feedback and it is used
to keep tuples from vacuuming as well as catalog_xmin for decoding stuff. Not
sure, xmin is applicable for logical replication.

The confirmed flush lsn is used as a startpoint when a peer node doesn't provide
the start lsn and to check that the start lsn is not older than the latest
confirmed flush lsn. The saving of the slot on disk at each call of
LogicalConfirmReceivedLocation doesn't help to avoid conflicts completely, but
it helps to decrease the probability of conflicts. So, i'm still not sure, we
need to save logical slots on each advance to avoid conflicts, because it
doesn't help in general. The conflicts should be resolved by other means.

Once, we truncate old wal segments in checkpoint only. I believe, it is ok if we
calculate the oldest lsn only at the beginning of the checkpoint, as it was in
the alternative solution. I think, we can update xmin horizon in checkpoint only
but the horizon advancing will be more lazy in this case.

Taking into account these thoughts, I can't see any problems with the alternative
patch where oldest wal lsn is calculated only in checkpoint.

With best regards,
Vitaly

On Thu, Jun 5, 2025 at 8:51 PM Vitaly Davydov <v.davydov@postgrespro.ru> wrote:
>
> Dear Alexander, Amit
>
> Alexander Korotkov wrote:
> > Also, I've changed ReplicationSlotsComputeRequiredLSN() call to
> > CheckPointReplicationSlots() to update required LSN after
> > SaveSlotToPath() updated last_saved_restart_lsn.  This helps to pass
> > checks in 001_stream_rep.pl without additional hacks.
>
> Thank you for the improvement and patch preparation. I confirm the test is
> passed without additional hacks now.
>
> I sill do not understand why this solution is favored. It is, in my opinion,
> a non backward-compatible solution. In any case, I'm ok to go with this patch.
> If needed, I may prepare a backward-compatible solution where
> last_saved_restart_lsn values will be in an another place of the shmem, rather
> than in ReplicationSlot struct.
>

I think we can use this approach for HEAD and probably keep the
previous idea for backbranches. Keeping some value in shared_memory
per slot sounds risky to me in terms of introducing new bugs.

> I still would like to add my 5 cents to the discussion.
>
> The purpose of the xmin value is to prevent tuples from vacuuming. Slots'
> restart_lsn values are used to calculate the oldest lsn to keep WAL segments
> from removal in checkpoint. These processes are pretty independent.
>
> The logical slots are advanced in 2 steps. At the first step, the logical
> decoding stuff periodically sets consistent candidate values for catalog_xmin and
> restart_lsn. At the second step, when LogicalConfirmReceivedLocation is called,
> the candidate values are assigned on catalog_xmin and restart_lsn values based
> on the confirmed lsn value. The slot is saved with these consistent values.
> It is important, that the candidate values are consistent, decoding guarantees
> it. In case of a crash, we should guarantee that the loaded from the disk
> catalog_xmin and restart_lsn values are consistent and valid for logical slots.
> LogicalConfirmReceivedLocation function keeps this consistency by updating them
> from consistent candidate values in a single operation.
>
> We have to guarantee that we use saved to disk values to calculate xmin horizon
> and slots' oldest lsn. For this purpose, effective_catalog_xmin is used. We
> update effective_catalog_xmin in LogicalConfirmReceivedLocation just
> after saving slot to disk. Another place where we update effective_catalog_xmin
> is when walsender receives hot standby feedback message.
>
> Once, we have two independent processes (vacuuming, checkpoint), we can calculate
> xmin horizon and oldest WAL lsn values independently (at different times) from
> the saved to disk values. Note, these values are updated in a non atomic way.
>
> The xmin value is set when the node receives hot standby feedback and it is used
> to keep tuples from vacuuming as well as catalog_xmin for decoding stuff.
>

Yeah, but with physical slots, it is possible that the slot's xmin
value is pointing to some value, say 700 (after restart), but vacuum
would have removed tuples from transaction IDs greater than 700 as
explained in email [1].

 Not
> sure, xmin is applicable for logical replication.
>
> The confirmed flush lsn is used as a startpoint when a peer node doesn't provide
> the start lsn and to check that the start lsn is not older than the latest
> confirmed flush lsn. The saving of the slot on disk at each call of
> LogicalConfirmReceivedLocation doesn't help to avoid conflicts completely, but
> it helps to decrease the probability of conflicts.
>

We don't save slots at each call of LogicalConfirmReceivedLocation()
and when we save also, it is not to avoid conflicts but to avoid
removing required WAL segments and tuples.

>
 So, i'm still not sure, we
> need to save logical slots on each advance to avoid conflicts, because it
> doesn't help in general. The conflicts should be resolved by other means.
>
> Once, we truncate old wal segments in checkpoint only. I believe, it is ok if we
> calculate the oldest lsn only at the beginning of the checkpoint, as it was in
> the alternative solution. I think, we can update xmin horizon in checkpoint only
> but the horizon advancing will be more lazy in this case.
>
> Taking into account these thoughts, I can't see any problems with the alternative
> patch where oldest wal lsn is calculated only in checkpoint.
>

The alternative will needlessly prevent removing WAL segments in some
cases when logical slots are in use.

[1] - https://www.postgresql.org/message-id/CAA4eK1KMaPA5jir_SFu%2Bqr3qu55OOdFWVZpuUkqTSGZ9fyPpHA%40mail.gmail.com

--
With Regards,
Amit Kapila.

On Tue, Jun 3, 2025 at 6:51 PM Alexander Korotkov <aekorotkov@gmail.com> wrote:
>
> >
> > As per my understanding, for logical slots, effective_xmin is only set
> > during the initial copy phase (or say if one has to export a
> > snapshot), after that, its value won't change. Please read the
> > comments in CreateInitDecodingContext() where we set its value. If you
> > still have questions about it, we can discuss further.
>
> OK, thank you for the clarification.   I've read the comments in
> CreateInitDecodingContext() as you suggested.  All of above makes me
> think *_xmin fields are handled properly.
>

Yes, they handled properly for logical slots, but there is no similar
safety mechanism for physical slots.

One minor comment:
+
+ /* Latest restart_lsn that has been flushed to disk. For persistent slots
+ * the flushed LSN should be taken into account when calculating the oldest

This doesn't follow our practice for multi-line comments.

--
With Regards,
Amit Kapila.