Are ZFS snapshots unsafe when PGSQL is spreading through multiple zpools?

On Sun, Jan 15, 2023 at 8:18 PM HECTOR INGERTO <HECTOR_25E@hotmail.com>
wrote:

Hello everybody,

I’m using PostgreSQL on openZFS. I use ZFS snapshots as a backup +
hotspare method.

From man zfs-snapshot: “Snapshots are taken atomically, so that all
snapshots correspond to the same moment in time.” So if a PSQL instance is
started from a zfs snapshot, it will start to replay the WAL from the last
checkpoint, in the same way it would do in a crash or power loss scenario. So
from my knowledge, ZFS snapshots can be used to rollback to a previous
point in time. Also, sending those snapshots to other computers will allow
you to have hotspares and remote backups. If I’m wrong here, I would
appreciate being told about it because I’m basing the whole question on
this premise.

On the other hand, we have the tablespace PGSQL feature, which is great
because it allows “unimportant” big data to be written into cheap HDD and
frequently used data into fast NVMe.

So far, so good. The problem is when both ideas are merged. Then,
snapshots from different pools are NOT atomical, snapshot on the HDD pool
isn’t going to be done at the same exact time as the one on the SSD pool,
and I don’t know enough about PGSQL internals to know how dangerous this
is. So here is where I would like to ask for your help with the following
questions:

First of all, what kind of problem can this lead to? Are we talking about
potential whole DB corruption or only the loss of a few of the latest
transactions?

Silent data corruption. *not* just losing your latest transaction.

In second place, if I’m initializing a corrupted PGSQL instance because
ZFS snapshots are from different pools and slightly different times, am I
going to notice it somehow or is it going to fail silently?

Silent. You might notice at the application level. Might.

In third and last place, is there some way to quantify the amount of risk
taken when snapshotting a PGSQL instance spread across two (or more)
different pools?

"Don't do it".

If you can't get atomic snapshots, don't do it, period.

You can use them together with a regular online backup. That is
pg_start_backup() // <snapshot multiple volumes> // pg_stop_backup()
together with log archiving. That's a perfectly valid method. But you
cannot and should not rely on snapshots alone.

--
Magnus Hagander
Me: https://www.hagander.net/ <http://www.hagander.net/&gt;
Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/&gt;

But you cannot and should not rely on snapshots alone

That’s only for non atomic (multiple pools) snapshots. Isn’t?

If I need to rely only on ZFS (automated) snapshots, then the best option would be to have two DB? Each one in each own pool. One HDD DB and one SSD DB. Then, the backend code should know on which DB the requested data is.

De: Magnus Hagander<mailto:magnus@hagander.net>
Enviado: domingo, 15 de enero de 2023 20:36
Para: HECTOR INGERTO<mailto:HECTOR_25E@hotmail.com>
CC: pgsql-general@postgresql.org<mailto:pgsql-general@postgresql.org>
Asunto: Re: Are ZFS snapshots unsafe when PGSQL is spreading through multiple zpools?

On Sun, Jan 15, 2023 at 8:18 PM HECTOR INGERTO <HECTOR_25E@hotmail.com<mailto:HECTOR_25E@hotmail.com>> wrote:
Hello everybody,

I’m using PostgreSQL on openZFS. I use ZFS snapshots as a backup + hotspare method.

From man zfs-snapshot: “Snapshots are taken atomically, so that all snapshots correspond to the same moment in time.” So if a PSQL instance is started from a zfs snapshot, it will start to replay the WAL from the last checkpoint, in the same way it would do in a crash or power loss scenario. So from my knowledge, ZFS snapshots can be used to rollback to a previous point in time. Also, sending those snapshots to other computers will allow you to have hotspares and remote backups. If I’m wrong here, I would appreciate being told about it because I’m basing the whole question on this premise.

On the other hand, we have the tablespace PGSQL feature, which is great because it allows “unimportant” big data to be written into cheap HDD and frequently used data into fast NVMe.

So far, so good. The problem is when both ideas are merged. Then, snapshots from different pools are NOT atomical, snapshot on the HDD pool isn’t going to be done at the same exact time as the one on the SSD pool, and I don’t know enough about PGSQL internals to know how dangerous this is. So here is where I would like to ask for your help with the following questions:

First of all, what kind of problem can this lead to? Are we talking about potential whole DB corruption or only the loss of a few of the latest transactions?

Silent data corruption. *not* just losing your latest transaction.

In second place, if I’m initializing a corrupted PGSQL instance because ZFS snapshots are from different pools and slightly different times, am I going to notice it somehow or is it going to fail silently?

Silent. You might notice at the application level. Might.

In third and last place, is there some way to quantify the amount of risk taken when snapshotting a PGSQL instance spread across two (or more) different pools?

"Don't do it".

If you can't get atomic snapshots, don't do it, period.

You can use them together with a regular online backup. That is pg_start_backup() // <snapshot multiple volumes> // pg_stop_backup() together with log archiving. That's a perfectly valid method. But you cannot and should not rely on snapshots alone.

--
Magnus Hagander
Me: https://www.hagander.net/&lt;https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.hagander.net%2F&amp;data=05%7C01%7C%7C4860509b67ea484420fb08daf72fddd4%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C638094082195595508%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=LRa%2BFTXpoZNsMLMrNLbL6xmgo9I3Mxx2CcCAh6nmguU%3D&amp;reserved=0&gt;
Work: https://www.redpill-linpro.com/&lt;https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.redpill-linpro.com%2F&amp;data=05%7C01%7C%7C4860509b67ea484420fb08daf72fddd4%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C638094082195752157%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=ziYhcTa5YOvHZEr2xk2nEKvSjLICE75zRhhCehvzIMs%3D&amp;reserved=0&gt;

On Sun, Jan 15, 2023 at 10:57 PM HECTOR INGERTO <hector_25e@hotmail.com>
wrote:

But you cannot and should not rely on snapshots alone

That’s only for non atomic (multiple pools) snapshots. Isn’t?

Right. For single-filesystem installs it should be fine. Just make sure it
has both the data and the WAL directories in the same one.

If I need to rely only on ZFS (automated) snapshots, then the best option
would be to have two DB? Each one in each own pool. One HDD DB and one SSD
DB. Then, the backend code should know on which DB the requested data is.

You could. I wouldn't -- I would set it up to use proper backups instead,
maybe using snapshots as the infrastructure. That gives you other
advantages as well, like being able to do PITR. It's a little more to set
up, but I'd say it's worth it.

//Magnus

We regularly use "history" databases.  Put them on slow media, and only take
a backup when data is added to them (monthly, quarterly, etc).

On 1/15/23 15:57, HECTOR INGERTO wrote:

But you cannot and should not rely on snapshots alone

That’s only for non atomic (multiple pools) snapshots. Isn’t?

If I need to rely only on ZFS (automated) snapshots, then the best option
would be to have two DB? Each one in each own pool. One HDD DB and one SSD
DB. Then, the backend code should know on which DB the requested data is.

*De: *Magnus Hagander <mailto:magnus@hagander.net>
*Enviado: *domingo, 15 de enero de 2023 20:36
*Para: *HECTOR INGERTO <mailto:HECTOR_25E@hotmail.com>
*CC: *pgsql-general@postgresql.org
*Asunto: *Re: Are ZFS snapshots unsafe when PGSQL is spreading through
multiple zpools?

On Sun, Jan 15, 2023 at 8:18 PM HECTOR INGERTO <HECTOR_25E@hotmail.com> wrote:

Hello everybody,

I’m using PostgreSQL on openZFS. I use ZFS snapshots as a backup +
hotspare method.

From man zfs-snapshot: “Snapshots are taken atomically, so that all
snapshots correspond to the same moment in time.” So if a PSQL
instance is started from a zfs snapshot, it will start to replay the
WAL from the last checkpoint, in the same way it would do in a crash
or power loss scenario. So from my knowledge, ZFS snapshots can be
used to rollback to a previous point in time. Also, sending those
snapshots to other computers will allow you to have hotspares and
remote backups. If I’m wrong here, I would appreciate being told about
it because I’m basing the whole question on this premise.

On the other hand, we have the tablespace PGSQL feature, which is
great because it allows “unimportant” big data to be written into
cheap HDD and frequently used data into fast NVMe.

So far, so good. The problem is when both ideas are merged. Then,
snapshots from different pools are NOT atomical, snapshot on the HDD
pool isn’t going to be done at the same exact time as the one on the
SSD pool, and I don’t know enough about PGSQL internals to know how
dangerous this is. So here is where I would like to ask for your help
with the following questions:

First of all, what kind of problem can this lead to? Are we talking
about potential whole DB corruption or only the loss of a few of the
latest transactions?

Silent data corruption. *not* just losing your latest transaction.

In second place, if I’m initializing a corrupted PGSQL instance
because ZFS snapshots are from different pools and slightly different
times, am I going to notice it somehow or is it going to fail silently?

Silent. You might notice at the application level. Might.

In third and last place, is there some way to quantify the amount of
risk taken when snapshotting a PGSQL instance spread across two (or
more) different pools?

"Don't do it".

If you can't get atomic snapshots, don't do it, period.

You can use them together with a regular online backup. That is
pg_start_backup() // <snapshot multiple volumes> // pg_stop_backup()
together with log archiving. That's a perfectly valid method. But you
cannot and should not rely on snapshots alone.

--

 Magnus Hagander
 Me: https://www.hagander.net/
<https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.hagander.net%2F&amp;data=05%7C01%7C%7C4860509b67ea484420fb08daf72fddd4%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C638094082195595508%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=LRa%2BFTXpoZNsMLMrNLbL6xmgo9I3Mxx2CcCAh6nmguU%3D&amp;reserved=0&gt;
 Work: https://www.redpill-linpro.com/
<https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.redpill-linpro.com%2F&amp;data=05%7C01%7C%7C4860509b67ea484420fb08daf72fddd4%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C638094082195752157%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=ziYhcTa5YOvHZEr2xk2nEKvSjLICE75zRhhCehvzIMs%3D&amp;reserved=0&gt;

--
Born in Arizona, moved to Babylonia.

I have understood I shall not do it, but could the technical details be discussed about why silent DB corruption can occur with non-atomical snapshots?

On Mon, 2023-01-16 at 08:41 +0000, HECTOR INGERTO wrote:

I have understood I shall not do it, but could the technical details be discussed about
why silent DB corruption can occur with non-atomical snapshots?

The database relies on the data being consistent when it performs crash recovery.
Imagine that a checkpoint is running while you take your snapshot. The checkpoint
syncs a data file with a new row to disk. Then it writes a WAL record and updates
the control file. Now imagine that the table with the new row is on a different
file system, and your snapshot captures the WAL and the control file, but not
the new row (it was still sitting in the kernel page cache when the snapshot was taken).
You end up with a lost row.

That is only one scenario. Many other ways of corruption can happen.

Yours,
Laurenz Albe
--
Cybertec | https://www.cybertec-postgresql.com

The database relies on the data being consistent when it performs crash recovery.
Imagine that a checkpoint is running while you take your snapshot. The checkpoint
syncs a data file with a new row to disk. Then it writes a WAL record and updates
the control file. Now imagine that the table with the new row is on a different
file system, and your snapshot captures the WAL and the control file, but not
the new row (it was still sitting in the kernel page cache when the snapshot was taken).
You end up with a lost row.

That is only one scenario. Many other ways of corruption can happen.

Can we say then that the risk comes only from the possibility of a checkpoint running inside the time gap between the non-simultaneous snapshots?

On 16 Jan 2023, at 15:37, HECTOR INGERTO <HECTOR_25E@hotmail.com> wrote:

The database relies on the data being consistent when it performs crash recovery.
Imagine that a checkpoint is running while you take your snapshot. The checkpoint
syncs a data file with a new row to disk. Then it writes a WAL record and updates
the control file. Now imagine that the table with the new row is on a different
file system, and your snapshot captures the WAL and the control file, but not
the new row (it was still sitting in the kernel page cache when the snapshot was taken).
You end up with a lost row.

That is only one scenario. Many other ways of corruption can happen.

Can we say then that the risk comes only from the possibility of a checkpoint running inside the time gap between the non-simultaneous snapshots?

I recently followed a course on distributed algorithms and recognised one of the patterns here.

The problem boils down to a distributed snapshotting algorithm, where both ZFS filesystem processes each initiate their own snapshot independently.

Without communicating with each other and with the database which messages (in this case traffic to and from the database to each FS) are part of their snapshots (sent or received), there are chances of lost messages, where either none of the process snapshots know that a 'message' was sent or none received it.

Algorithms like Tarry, Lai-Yang or the Echo algorithm solve this by adding communication between those processes about messages in transit.

Alban Hertroys
--
There is always an exception to always.

On Mon, Jan 16, 2023 at 3:37 PM HECTOR INGERTO <hector_25e@hotmail.com>
wrote:

The database relies on the data being consistent when it performs crash

recovery.

Imagine that a checkpoint is running while you take your snapshot. The

checkpoint

syncs a data file with a new row to disk. Then it writes a WAL record

and updates

the control file. Now imagine that the table with the new row is on a

different

file system, and your snapshot captures the WAL and the control file,

but not

the new row (it was still sitting in the kernel page cache when the

snapshot was taken).

You end up with a lost row.

That is only one scenario. Many other ways of corruption can happen.

Can we say then that the risk comes only from the possibility of a
checkpoint running inside the time gap between the non-simultaneous
snapshots?

No. The risk comes from any write at all, foreground or background,
occurring inside this gap. Checkpoints are only one of many types of writes
that can cause the problems.

The data integrity is dependent on writes happening in the order they are
issued/confirmed when they are synchronous. Thus terms like WAL = Write
*Ahead* Log.

--
Magnus Hagander
Me: https://www.hagander.net/ <http://www.hagander.net/&gt;
Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/&gt;

On Mon, 2023-01-16 at 14:37 +0000, HECTOR INGERTO wrote:

The database relies on the data being consistent when it performs crash recovery.
Imagine that a checkpoint is running while you take your snapshot.  The checkpoint
syncs a data file with a new row to disk.  Then it writes a WAL record and updates
the control file.  Now imagine that the table with the new row is on a different
file system, and your snapshot captures the WAL and the control file, but not
the new row (it was still sitting in the kernel page cache when the snapshot was taken).
You end up with a lost row.

That is only one scenario.  Many other ways of corruption can happen.

 
Can we say then that the risk comes only from the possibility of a checkpoint running
inside the time gap between the non-simultaneous snapshots?

Another case: a transaction COMMITs, and a slightly later transaction reads the data
and sets a hint bit. If the snapshot of the file system with the data directory in it
is slightly later than the snapshot of the file system with "pg_wal", the COMMIT might
not be part of the snapshot, but the hint bit could be.

Then these uncommitted data could be visible if you recover from the snapshot.

Yours,
Laurenz Albe
--
Cybertec | https://www.cybertec-postgresql.com

Another case: a transaction COMMITs, and a slightly later transaction reads the data
and sets a hint bit. If the snapshot of the file system with the data directory in it
is slightly later than the snapshot of the file system with "pg_wal", the COMMIT might
not be part of the snapshot, but the hint bit could be.

Then these uncommitted data could be visible if you recover from the snapshot.

Yours,
Laurenz Albe

Thank you all. I have it clearer now.

As a last point. Making the snapshot to the WAL dataset first or last would make any difference?

On Tue, Jan 17, 2023 at 03:22:02PM +0000, HECTOR INGERTO wrote:

Another case: a transaction COMMITs, and a slightly later transaction reads

the data

and sets a hint bit. If the snapshot of the file system with the data

directory in it

is slightly later than the snapshot of the file system with "pg_wal", the

COMMIT might

not be part of the snapshot, but the hint bit could be.

Then these uncommitted data could be visible if you recover from the

snapshot.

Yours,
Laurenz Albe

Thank you all. I have it clearer now.

As a last point. Making the snapshot to the WAL dataset first or last would
make any difference?

How would you know which WAL snapshot to use, and if there is any
missing data between them?

I have often wondered if we could document rules where multiple
asychronous snapshots would be safe but have never gotten very far.

--
Bruce Momjian <bruce@momjian.us> https://momjian.us
EDB https://enterprisedb.com

Embrace your flaws. They make you human, rather than perfect,
which you will never be.

On Tue, 2023-01-17 at 15:22 +0000, HECTOR INGERTO wrote:

Another case: a transaction COMMITs, and a slightly later transaction reads the data
and sets a hint bit.  If the snapshot of the file system with the data directory in it
is slightly later than the snapshot of the file system with "pg_wal", the COMMIT might
not be part of the snapshot, but the hint bit could be.

Then these uncommitted data could be visible if you recover from the snapshot.

 
Thank you all. I have it clearer now.
 
As a last point. Making the snapshot to the WAL dataset first or last would make any difference?

Imagine you run DROP TABLE. During the implicit COMMIT at the end of the statement,
the files behind the table are deleted. If the snapshot of "pg_wal" is earlier than
the snapshot of the data files, you end up with a table that is not yet dropped,
but the files are gone.

I won't try to find an example if you now ask what if no checkpoint ends between the
statements, the snapshot on "pg_wal" is earlier and we don't run DROP TABLE.

Why do you go to all this effort rather than performing a correct backup?

Yours,
Laurenz Albe

I wanted to understand the underlying issue.

I use ZFS snapshots instead of a “correct” backup because with only two machines it allows me to have backups in the main machine and in the secondary too that acts as hotspare at the same time.

To accomplish the same I would need 3 nodes. The main, the replica hotspare and the proper backup.

De: Laurenz Albe<mailto:laurenz.albe@cybertec.at>
Enviado: miércoles, 18 de enero de 2023 11:02
Para: HECTOR INGERTO<mailto:hector_25e@hotmail.com>; Magnus Hagander<mailto:magnus@hagander.net>
CC: pgsql-general@postgresql.org <pgsql-general@postgresql.org><mailto:pgsql-general@postgresql.org>
Asunto: Re: Are ZFS snapshots unsafe when PGSQL is spreading through multiple zpools?

On Tue, 2023-01-17 at 15:22 +0000, HECTOR INGERTO wrote:

Another case: a transaction COMMITs, and a slightly later transaction reads the data
and sets a hint bit. If the snapshot of the file system with the data directory in it
is slightly later than the snapshot of the file system with "pg_wal", the COMMIT might
not be part of the snapshot, but the hint bit could be.

Then these uncommitted data could be visible if you recover from the snapshot.

Thank you all. I have it clearer now.

As a last point. Making the snapshot to the WAL dataset first or last would make any difference?

Imagine you run DROP TABLE. During the implicit COMMIT at the end of the statement,
the files behind the table are deleted. If the snapshot of "pg_wal" is earlier than
the snapshot of the data files, you end up with a table that is not yet dropped,
but the files are gone.

I won't try to find an example if you now ask what if no checkpoint ends between the
statements, the snapshot on "pg_wal" is earlier and we don't run DROP TABLE.

Why do you go to all this effort rather than performing a correct backup?

Yours,
Laurenz Albe

On 1/18/23 09:38, HECTOR INGERTO wrote:

I wanted to understand the underlying issue.

I use ZFS snapshots instead of a “correct” backup because with only
two machines it allows me to have backups in the main machine and in
the secondary too that acts as hotspare at the same time.

To accomplish the same I would need 3 nodes. The main, the replica
hotspare and the proper backup.

Isn't "proper backup" typically thought of as offsite reloadable file
capture, hopefully relatively recent?  Disaster recovery, and all that.

On 1/18/23 10:54, Rob Sargent wrote:

On 1/18/23 09:38, HECTOR INGERTO wrote:

I wanted to understand the underlying issue.

I use ZFS snapshots instead of a “correct” backup because with only two
machines it allows me to have backups in the main machine and in the
secondary too that acts as hotspare at the same time.

To accomplish the same I would need 3 nodes. The main, the replica
hotspare and the proper backup.

Isn't "proper backup" typically thought of as offsite reloadable file
capture, hopefully relatively recent?  Disaster recovery, and all that.

That is in fact what "proper backup" means.  "Hot spare" means high
availability, and *HA is not DR*.

--
Born in Arizona, moved to Babylonia.