Thread: Revisiting {CREATE INDEX, REINDEX} CONCURRENTLY improvements

Hello, hackers!

I think about revisiting (1) ({CREATE INDEX, REINDEX} CONCURRENTLY
improvements) in some lighter way.

Yes, a serious bug was (2) caused by this optimization and now it reverted.

But what about a more safe idea in that direction:
1) add new horizon which ignores PROC_IN_SAFE_IC backends and standbys queries
2) use this horizon for settings LP_DEAD bit in indexes (excluding
indexes being built of course)

Index LP_DEAD hints are not used by standby in any way (they are just
ignored), also heap scan done by index building does not use them as
well.

But, at the same time:
1) index scans will be much faster during index creation or standby
reporting queries
2) indexes can keep them fit using different optimizations
3) less WAL due to a huge amount of full pages writes (which caused by
tons of LP_DEAD in indexes)

The patch seems more-less easy to implement.
Does it worth being implemented? Or to scary?

[1]: https://postgr.es/m/20210115133858.GA18931@alvherre.pgsql
[2]: https://postgr.es/m/17485-396609c6925b982d%40postgresql.org

Hello!

> I've thought about alternative solutions, too: how about getting a new snapshot every so often?
> We don't really care about the liveness of the already-scanned data; the snapshots used for RIC
> are used only during the scan. C/RIC's relation's lock level means vacuum can't run to clean up
> dead line items, so as long as we only swap the backend's reported snapshot (thus xmin) while
> the scan is between pages we should be able to reduce the time C/RIC is the one backend
> holding back cleanup of old tuples.

Hm, it looks like an interesting idea! It may be more dangerous, but
at least it feels much more elegant than an LP_DEAD-related way.
Also, feels like we may apply this to both phases (first and the second scans).
The original patch (1) was helping only to the second one (after call
to set_indexsafe_procflags).

But for the first scan we allowed to do so only for non-unique indexes
because of:

> * The reason for doing that is to avoid
> * bogus unique-index failures due to concurrent UPDATEs (we might see
> * different versions of the same row as being valid when we pass over them,
> * if we used HeapTupleSatisfiesVacuum).  This leaves us with an index that
> * does not contain any tuples added to the table while we built the index.

Also, (1) was limited to indexes without expressions and predicates
(2) because such may execute queries to other tables (sic!).
One possible solution is to add some checks to make sure no
user-defined functions are used.
But as far as I understand, it affects only CIC for now and does not
affect the ability to use the proposed technique (updating snapshot
time to time).

However, I think we need some more-less formal proof it is safe - it
is really challenging to keep all the possible cases in the head. I’ll
try to do something here.
Another possible issue may be caused by the new locking pattern - we
will be required to wait for all transaction started before the ending
of the phase to exit.

[1]: https://postgr.es/m/20210115133858.GA18931@alvherre.pgsql
[2]:
https://www.postgresql.org/message-id/flat/CAAaqYe_tq_Mtd9tdeGDsgQh%2BwMvouithAmcOXvCbLaH2PPGHvA%40mail.gmail.com#cbe3997b75c189c3713f243e25121c20

On Sun, 17 Dec 2023, 21:14 Michail Nikolaev, <michail.nikolaev@gmail.com> wrote:
>
> Hello!
>
> > I've thought about alternative solutions, too: how about getting a new snapshot every so often?
> > We don't really care about the liveness of the already-scanned data; the snapshots used for RIC
> > are used only during the scan. C/RIC's relation's lock level means vacuum can't run to clean up
> > dead line items, so as long as we only swap the backend's reported snapshot (thus xmin) while
> > the scan is between pages we should be able to reduce the time C/RIC is the one backend
> > holding back cleanup of old tuples.
>
> Hm, it looks like an interesting idea! It may be more dangerous, but
> at least it feels much more elegant than an LP_DEAD-related way.
> Also, feels like we may apply this to both phases (first and the second scans).
> The original patch (1) was helping only to the second one (after call
> to set_indexsafe_procflags).
>
> But for the first scan we allowed to do so only for non-unique indexes
> because of:
>
> > * The reason for doing that is to avoid
> > * bogus unique-index failures due to concurrent UPDATEs (we might see
> > * different versions of the same row as being valid when we pass over them,
> > * if we used HeapTupleSatisfiesVacuum).  This leaves us with an index that
> > * does not contain any tuples added to the table while we built the index.

Yes, for that we'd need an extra scan of the index that validates
uniqueness. I think there was a proposal (though it may only have been
an idea) some time ago, about turning existing non-unique indexes into
unique ones by validating the data. Such a system would likely be very
useful to enable this optimization.

> Also, (1) was limited to indexes without expressions and predicates
> (2) because such may execute queries to other tables (sic!).

Note that the use of such expressions would be a violation of the
function's definition; it would depend on data from other tables which
makes the function behave like a STABLE function, as opposed to the
IMMUTABLE that is required for index expressions. So, I don't think we
should specially care about being correct for incorrectly marked
function definitions.

> One possible solution is to add some checks to make sure no
> user-defined functions are used.
> But as far as I understand, it affects only CIC for now and does not
> affect the ability to use the proposed technique (updating snapshot
> time to time).
>
> However, I think we need some more-less formal proof it is safe - it
> is really challenging to keep all the possible cases in the head. I’ll
> try to do something here.

I just realised there is one issue with this design: We can't cheaply
reset the snapshot during the second table scan:
It is critically important that the second scan of R/CIC uses an index
contents summary (made with index_bulk_delete) that was created while
the current snapshot was already registered. If we didn't do that, the
following would occur:

1. The index is marked ready for inserts from concurrent backends, but
not yet ready for queries.
2. We get the bulkdelete data
3. A concurrent backend inserts a new tuple T on heap page P, inserts
it into the index, and commits. This tuple is not in the summary, but
has been inserted into the index.
4. R/CIC resets the snapshot, making T visible.
5. R/CIC scans page P, finds that tuple T has to be indexed but is not
present in the summary, thus inserts that tuple into the index (which
already had it inserted at 3)

This thus would be a logic bug, as indexes assume at-most-once
semantics for index tuple insertion; duplicate insertion are an error.

So, the "reset the snapshot every so often" trick cannot be applied in
phase 3 (the rescan), or we'd have to do an index_bulk_delete call
every time we reset the snapshot. Rescanning might be worth the cost
(e.g. when using BRIN), but that is very unlikely.

Alternatively, we'd need to find another way to prevent us from
inserting these duplicate entries - maybe by storing the scan's data
in a buffer to later load into the index after another
index_bulk_delete()? Counterpoint: for BRIN indexes that'd likely
require a buffer much larger than the result index would be.

Either way, for the first scan (i.e. phase 2 "build new indexes") this
is not an issue: we don't care about what transaction adds/deletes
tuples at that point.
For all we know, all tuples of the table may be deleted concurrently
before we even allow concurrent backends to start inserting tuples,
and the algorithm would still work as it does right now.

> Another possible issue may be caused by the new locking pattern - we
> will be required to wait for all transaction started before the ending
> of the phase to exit.

What do you mean by "new locking pattern"? We already keep an
ShareUpdateExclusiveLock on every heap table we're accessing during
R/CIC, and that should already prevent any concurrent VACUUM
operations, right?

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

Hello!

> Also, feels like we may apply this to both phases (first and the second scans).
> The original patch (1) was helping only to the second one (after call
> to set_indexsafe_procflags).

Oops, I was wrong here. The original version of the patch was also applied to
both phases.

> Note that the use of such expressions would be a violation of the
> function's definition; it would depend on data from other tables which
> makes the function behave like a STABLE function, as opposed to the
> IMMUTABLE that is required for index expressions. So, I don't think we
> should specially care about being correct for incorrectly marked
> function definitions.

Yes, but such cases could probably cause crashes also...
So, I think it is better to check them for custom functions. But I
still not sure -
if such limitations still required for proposed optimization or not.

> I just realised there is one issue with this design: We can't cheaply
> reset the snapshot during the second table scan:
> It is critically important that the second scan of R/CIC uses an index
> contents summary (made with index_bulk_delete) that was created while
> the current snapshot was already registered.

> So, the "reset the snapshot every so often" trick cannot be applied in
> phase 3 (the rescan), or we'd have to do an index_bulk_delete call
> every time we reset the snapshot. Rescanning might be worth the cost
> (e.g. when using BRIN), but that is very unlikely.

Hm, I think it is still possible. We could just manually recheck the
tuples we see
to the snapshot currently used for the scan. If an "old" snapshot can see
the tuple also (HeapTupleSatisfiesHistoricMVCC) then search for it in the
index summary.

> What do you mean by "new locking pattern"? We already keep an
> ShareUpdateExclusiveLock on every heap table we're accessing during
> R/CIC, and that should already prevent any concurrent VACUUM
> operations, right?

I was thinking not about "classical" locking, but about waiting for
other backends
by WaitForLockers(heaplocktag, ShareLock, true). But I think
everything should be
fine.

Best regards,
Michail.

On Wed, 20 Dec 2023 at 10:56, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
> > Note that the use of such expressions would be a violation of the
> > function's definition; it would depend on data from other tables which
> > makes the function behave like a STABLE function, as opposed to the
> > IMMUTABLE that is required for index expressions. So, I don't think we
> > should specially care about being correct for incorrectly marked
> > function definitions.
>
> Yes, but such cases could probably cause crashes also...
> So, I think it is better to check them for custom functions. But I
> still not sure -
> if such limitations still required for proposed optimization or not.

I think contents could be inconsistent, but not more inconsistent than
if the index was filled across multiple transactions using inserts.
Either way I don't see it breaking more things that are not already
broken in that way in other places - at most it will introduce another
path that exposes the broken state caused by mislabeled functions.

> > I just realised there is one issue with this design: We can't cheaply
> > reset the snapshot during the second table scan:
> > It is critically important that the second scan of R/CIC uses an index
> > contents summary (made with index_bulk_delete) that was created while
> > the current snapshot was already registered.
>
> > So, the "reset the snapshot every so often" trick cannot be applied in
> > phase 3 (the rescan), or we'd have to do an index_bulk_delete call
> > every time we reset the snapshot. Rescanning might be worth the cost
> > (e.g. when using BRIN), but that is very unlikely.
>
> Hm, I think it is still possible. We could just manually recheck the
> tuples we see
> to the snapshot currently used for the scan. If an "old" snapshot can see
> the tuple also (HeapTupleSatisfiesHistoricMVCC) then search for it in the
> index summary.
That's an interesting method.

How would this deal with tuples not visible to the old snapshot?
Presumably we can assume they're newer than that snapshot (the old
snapshot didn't have it, but the new one does, so it's committed after
the old snapshot, making them newer), so that backend must have
inserted it into the index already, right?

> HeapTupleSatisfiesHistoricMVCC

That function has this comment marker:
   "Only usable on tuples from catalog tables!"
Is that correct even for this?

Should this deal with any potential XID wraparound, too?
How does this behave when the newly inserted tuple's xmin gets frozen?
This would be allowed to happen during heap page pruning, afaik - no
rules that I know of which are against that - but it would create
issues where normal snapshot visibility rules would indicate it
visible to both snapshots regardless of whether it actually was
visible to the older snapshot when that snapshot was created...

Either way, "Historic snapshot" isn't something I've worked with
before, so that goes onto my "figure out how it works" pile.

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

Hello!

> How would this deal with tuples not visible to the old snapshot?
> Presumably we can assume they're newer than that snapshot (the old
> snapshot didn't have it, but the new one does, so it's committed after
> the old snapshot, making them newer), so that backend must have
> inserted it into the index already, right?

Yes, exactly.

>> HeapTupleSatisfiesHistoricMVCC
> That function has this comment marker:
> "Only usable on tuples from catalog tables!"
> Is that correct even for this?

Yeah, we just need HeapTupleSatisfiesVisibility (which calls
HeapTupleSatisfiesMVCC) instead.

> Should this deal with any potential XID wraparound, too?

Yeah, looks like we should care about such case somehow.

Possible options here:

1) Skip vac_truncate_clog while CIC is running. In fact, I think it's
not that much worse than the current state - datfrozenxid is still
updated in the catalog and will be considered the next time
vac_update_datfrozenxid is called (the next VACCUM on any table).

2) Delay vac_truncate_clog while CIC is running.
In such a case, if it was skipped, we will need to re-run it using the
index builds backend later.

3) Wait for 64-bit xids :)

4) Any ideas?

In addition, for the first and second options, we need logic to cancel
the second phase in the case of ForceTransactionIdLimitUpdate.
But maybe I'm missing something and the tuples may be frozen, ignoring
the set datfrozenxid values (over some horizon calculated at runtime
based on the xmin backends).

> How does this behave when the newly inserted tuple's xmin gets frozen?
> This would be allowed to happen during heap page pruning, afaik - no
> rules that I know of which are against that - but it would create
> issues where normal snapshot visibility rules would indicate it
> visible to both snapshots regardless of whether it actually was
> visible to the older snapshot when that snapshot was created...

Yes, good catch.
Assuming we have somehow prevented vac_truncate_clog from occurring
during CIC, we can leave frozen and potentially frozen
(xmin<frozenXID) for the second phase.

So, first phase processing items:
* not frozen
* xmin>frozenXID (may not be frozen)
* visible by snapshot

second phase:
* frozen
* xmin>frozenXID (may be frozen)
* not in the index summary
* visible by "old" snapshot

You might also think – why is the first stage needed at all? Just use
batch processing during initial index building?

Best regards,
Mikhail.

> Yes, good catch.
> Assuming we have somehow prevented vac_truncate_clog from occurring
> during CIC, we can leave frozen and potentially frozen
> (xmin<frozenXID) for the second phase.

Just realized that we can leave this for the first stage to improve efficiency.
Since the ID is locked, anything that can be frozen will be visible in
the first stage.

Hello.

Realized my last idea is invalid (because tuples are frozen by using
dynamically calculated horizon) - so, don't waste your time on it :)

Need to think a little bit more here.

Thanks,
Mikhail.

Hello!

It seems like the idea of "old" snapshot is still a valid one.

> Should this deal with any potential XID wraparound, too?

As far as I understand in our case, we are not affected by this in any way.
Vacuum in our table is not possible because of locking, so, nothing
may be frozen (see below).
In the case of super long index building, transactional limits will
stop new connections using current
regular infrastructure because it is based on relation data (but not
actual xmin of backends).

> How does this behave when the newly inserted tuple's xmin gets frozen?
> This would be allowed to happen during heap page pruning, afaik - no
> rules that I know of which are against that - but it would create
> issues where normal snapshot visibility rules would indicate it
> visible to both snapshots regardless of whether it actually was
> visible to the older snapshot when that snapshot was created...

As I can see, heap_page_prune never freezes any tuples.
In the case of regular vacuum, it used this way: call heap_page_prune
and then call heap_prepare_freeze_tuple and then
heap_freeze_execute_prepared.

Merry Christmas,
Mikhail.

On Mon, 25 Dec 2023 at 15:12, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> Hello!
>
> It seems like the idea of "old" snapshot is still a valid one.
>
> > Should this deal with any potential XID wraparound, too?
>
> As far as I understand in our case, we are not affected by this in any way.
> Vacuum in our table is not possible because of locking, so, nothing
> may be frozen (see below).
> In the case of super long index building, transactional limits will
> stop new connections using current
> regular infrastructure because it is based on relation data (but not
> actual xmin of backends).
>
> > How does this behave when the newly inserted tuple's xmin gets frozen?
> > This would be allowed to happen during heap page pruning, afaik - no
> > rules that I know of which are against that - but it would create
> > issues where normal snapshot visibility rules would indicate it
> > visible to both snapshots regardless of whether it actually was
> > visible to the older snapshot when that snapshot was created...
>
> As I can see, heap_page_prune never freezes any tuples.
> In the case of regular vacuum, it used this way: call heap_page_prune
> and then call heap_prepare_freeze_tuple and then
> heap_freeze_execute_prepared.

Correct, but there are changes being discussed where we would freeze
tuples during pruning as well [0], which would invalidate that
implementation detail. And, if I had to choose between improved
opportunistic freezing and improved R/CIC, I'd probably choose
improved freezing over R/CIC.

As an alternative, we _could_ keep track of concurrent index inserts
using a dummy index (with the same predicate) which only holds the
TIDs of the inserted tuples. We'd keep it as an empty index in phase
1, and every time we reset the visibility snapshot we now only need to
scan that index to know what tuples were concurrently inserted. This
should have a significantly lower IO overhead than repeated full index
bulkdelete scans for the new index in the second table scan phase of
R/CIC. However, in a worst case it could still require another
O(tablesize) of storage.

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

[0] https://www.postgresql.org/message-id/CAAKRu_a+g2oe6aHJCbibFtNFiy2aib4E31X9QYJ_qKjxZmZQEg@mail.gmail.com

Hello!

> Correct, but there are changes being discussed where we would freeze
> tuples during pruning as well [0], which would invalidate that
> implementation detail. And, if I had to choose between improved
> opportunistic freezing and improved R/CIC, I'd probably choose
> improved freezing over R/CIC.

As another option, we could extract a dedicated horizon value for an
opportunistic freezing.
And use some flags in R/CIC backend to keep it at the required value.

Best regards,
Michail.

Hello, Melanie!

Sorry to interrupt you, just a quick question.

> Correct, but there are changes being discussed where we would freeze
> tuples during pruning as well [0], which would invalidate that
> implementation detail. And, if I had to choose between improved
> opportunistic freezing and improved R/CIC, I'd probably choose
> improved freezing over R/CIC.

Do you have any patches\threads related to that refactoring
(opportunistic freezing of tuples during pruning) [0]?
This may affect the idea of the current thread (latest version of it
mostly in [1]) - it may be required to disable such a feature for
particular relation temporary or affect horizon used for pruning
(without holding xmin).

Just no sure - is it reasonable to start coding right now, or wait for
some prune-freeze-related patch first?

[0] https://www.postgresql.org/message-id/CAAKRu_a+g2oe6aHJCbibFtNFiy2aib4E31X9QYJ_qKjxZmZQEg@mail.gmail.com
[1]
https://www.postgresql.org/message-id/flat/CANtu0ojRX%3DosoiXL9JJG6g6qOowXVbVYX%2BmDsN%2B2jmFVe%3DeG7w%40mail.gmail.com#a8ff53f23d0fc7edabd446b4d634e7b5

> > > I just realised there is one issue with this design: We can't cheaply
> > > reset the snapshot during the second table scan:
> > > It is critically important that the second scan of R/CIC uses an index
> > > contents summary (made with index_bulk_delete) that was created while
> > > the current snapshot was already registered.
> >
> > > So, the "reset the snapshot every so often" trick cannot be applied in
> > > phase 3 (the rescan), or we'd have to do an index_bulk_delete call
> > > every time we reset the snapshot. Rescanning might be worth the cost
> > > (e.g. when using BRIN), but that is very unlikely.
> >
> > Hm, I think it is still possible. We could just manually recheck the
> > tuples we see
> > to the snapshot currently used for the scan. If an "old" snapshot can see
> > the tuple also (HeapTupleSatisfiesHistoricMVCC) then search for it in the
> > index summary.
> That's an interesting method.
>
> How would this deal with tuples not visible to the old snapshot?
> Presumably we can assume they're newer than that snapshot (the old
> snapshot didn't have it, but the new one does, so it's committed after
> the old snapshot, making them newer), so that backend must have
> inserted it into the index already, right?

I made a draft of the patch and this idea is not working.

The problem is generally the same:

* reference snapshot sees tuple X
* reference snapshot is used to create index summary (but there is no
tuple X in the index summary)
* tuple X is updated to Y creating a HOT-chain
* we started scan with new temporary snapshot (it sees Y, X is too old for it)
* tuple X is pruned from HOT-chain because it is not protected by any snapshot
* we see tuple Y in the scan with temporary snapshot
    * it is not in the index summary - so, we need to check if
reference snapshot can see it
    * there is no way to understand if the reference snapshot was able
to see tuple X - because we need the full HOT chain (with X tuple) for
that

Best regards,
Michail.

On Thu, 1 Feb 2024, 17:06 Michail Nikolaev, <michail.nikolaev@gmail.com> wrote:
>
> > > > I just realised there is one issue with this design: We can't cheaply
> > > > reset the snapshot during the second table scan:
> > > > It is critically important that the second scan of R/CIC uses an index
> > > > contents summary (made with index_bulk_delete) that was created while
> > > > the current snapshot was already registered.

I think the best way for this to work would be an index method that
exclusively stores TIDs, and of which we can quickly determine new
tuples, too. I was thinking about something like GIN's format, but
using (generation number, tid) instead of ([colno, colvalue], tid) as
key data for the internal trees, and would be unlogged (because the
data wouldn't have to survive a crash). Then we could do something
like this for the second table scan phase:

0. index->indisready is set
[...]
1. Empty the "changelog index", resetting storage and the generation number.
2. Take index contents snapshot of new index, store this.
3. Loop until completion:
4a. Take visibility snapshot
4b. Update generation number of the changelog index, store this.
4c. Take index snapshot of "changelog index" for data up to the
current stored generation number. Not including, because we only need
to scan that part of the index that were added before we created our
visibility snapshot, i.e. TIDs labeled with generation numbers between
the previous iteration's generation number (incl.) and this
iteration's generation (excl.).
4d. Combine the current index snapshot with that of the "changelog"
index, and save this.
    Note that this needs to take care to remove duplicates.
4e. Scan segment of table (using the combined index snapshot) until we
need to update our visibility snapshot or have scanned the whole
table.

This should give similar, if not the same, behavour as that which we
have when we RIC a table with several small indexes, without requiring
us to scan a full index of data several times.

Attemp on proving this approach's correctness:
In phase 3, after each step 4b:
All matching tuples of the table that are in the visibility snapshot:
* Were created before scan 1's snapshot, thus in the new index's snapshot, or
* Were created after scan 1's snapshot but before index->indisready,
thus not in the new index's snapshot, nor in the changelog index, or
* Were created after the index was set as indisready, and committed
before the previous iteration's visibility snapshot, thus in the
combined index snapshot, or
* Were created after the index was set as indisready, after the
previous visibility snapshot was taken, but before the current
visibility snapshot was taken, and thus definitely included in the
changelog index.

Because we hold a snapshot, no data in the table that we should see is
removed, so we don't have a chance of broken HOT chains.


Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

Hello!

> I think the best way for this to work would be an index method that
> exclusively stores TIDs, and of which we can quickly determine new
> tuples, too. I was thinking about something like GIN's format, but
> using (generation number, tid) instead of ([colno, colvalue], tid) as
> key data for the internal trees, and would be unlogged (because the
> data wouldn't have to survive a crash)

Yeah, this seems to be a reasonable approach, but there are some
doubts related to it - it needs new index type as well as unlogged
indexes to be introduced - this may make the patch too invasive to be
merged. Also, some way to remove the index from the catalog in case of
a crash may be required.

A few more thoughts:
* it is possible to go without generation number - we may provide a
way to do some kind of fast index lookup (by TID) directly during the
second table scan phase.
* one more option is to maintain a Tuplesorts (instead of an index)
with TIDs as changelog and merge with index snapshot after taking a
new visibility snapshot. But it is not clear how to share the same
Tuplesort with multiple inserting backends.
* crazy idea - what is about to do the scan in the index we are
building? We have tuple, so, we have all the data indexed in the
index. We may try to do an index scan using that data to get all
tuples and find the one with our TID :) Yes, in some cases it may be
too bad because of the huge amount of TIDs we need to scan + also
btree copies whole page despite we need single item. But some
additional index method may help - feels like something related to
uniqueness (but it is only in btree anyway).

Thanks,
Mikhail.

One more idea - is just forbid HOT prune while the second phase is
running. It is not possible anyway currently because of snapshot held.

Possible enhancements:
* we may apply restriction only to particular tables
* we may apply restrictions only to part of the tables (not yet
scanned by R/CICs).

Yes, it is not an elegant solution, limited, not reliable in terms of
architecture, but a simple one.

On Wed, 21 Feb 2024 at 00:33, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> Hello!
> > I think the best way for this to work would be an index method that
> > exclusively stores TIDs, and of which we can quickly determine new
> > tuples, too. I was thinking about something like GIN's format, but
> > using (generation number, tid) instead of ([colno, colvalue], tid) as
> > key data for the internal trees, and would be unlogged (because the
> > data wouldn't have to survive a crash)
>
> Yeah, this seems to be a reasonable approach, but there are some
> doubts related to it - it needs new index type as well as unlogged
> indexes to be introduced - this may make the patch too invasive to be
> merged.

I suppose so, though persistence is usually just to keep things
correct in case of crashes, and this "index" is only there to support
processes that don't expect to survive crashes.

> Also, some way to remove the index from the catalog in case of
> a crash may be required.

That's less of an issue though, we already accept that a crash during
CIC/RIC leaves unusable indexes around, so "needs more cleanup" is not
exactly a blocker.

> A few more thoughts:
> * it is possible to go without generation number - we may provide a
> way to do some kind of fast index lookup (by TID) directly during the
> second table scan phase.

While possible, I don't think this would be more performant than the
combination approach, at the cost of potentially much more random IO
when the table is aggressively being updated.

> * one more option is to maintain a Tuplesorts (instead of an index)
> with TIDs as changelog and merge with index snapshot after taking a
> new visibility snapshot. But it is not clear how to share the same
> Tuplesort with multiple inserting backends.

Tuplesort requires the leader process to wait for concurrent backends
to finish their sort before it can start consuming their runs. This
would make it a very bad alternative to the "changelog index" as the
CIC process would require on-demand actions from concurrent backends
(flush of sort state). I'm not convinced that's somehow easier.

> * crazy idea - what is about to do the scan in the index we are
> building? We have tuple, so, we have all the data indexed in the
> index. We may try to do an index scan using that data to get all
> tuples and find the one with our TID :)

We can't rely on that, because we have no guarantee we can find the
tuple quickly enough. Equality-based indexing is very much optional,
and so are TID-based checks (outside the current vacuum-related APIs),
so finding one TID can (and probably will) take O(indexsize) when the
tuple is not in the index, which is one reason for ambulkdelete() to
exist.

Kind regards,

Matthias van de Meent

On Wed, 21 Feb 2024 at 09:35, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> One more idea - is just forbid HOT prune while the second phase is
> running. It is not possible anyway currently because of snapshot held.
>
> Possible enhancements:
> * we may apply restriction only to particular tables
> * we may apply restrictions only to part of the tables (not yet
> scanned by R/CICs).
>
> Yes, it is not an elegant solution, limited, not reliable in terms of
> architecture, but a simple one.

How do you suppose this would work differently from a long-lived
normal snapshot, which is how it works right now?
Would it be exclusively for that relation? How would this be
integrated with e.g. heap_page_prune_opt?

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

Hi!

> How do you suppose this would work differently from a long-lived
> normal snapshot, which is how it works right now?

Difference in the ability to take new visibility snapshot periodically
during the second phase with rechecking visibility of tuple according
to the "reference" snapshot (which is taken only once like now).
It is the approach from (1) but with a workaround for the issues
caused by heap_page_prune_opt.

> Would it be exclusively for that relation?
Yes, only for that affected relation. Other relations are unaffected.

> How would this be integrated with e.g. heap_page_prune_opt?
Probably by some flag in RelationData, but not sure here yet.

If the idea looks sane, I could try to extend my POC - it should be
not too hard, likely (I already have tests to make sure it is
correct).

(1):
https://www.postgresql.org/message-id/flat/CANtu0oijWPRGRpaRR_OvT2R5YALzscvcOTFh-%3DuZKUpNJmuZtw%40mail.gmail.com#8141eb2ea177ff560ee713b3f20de404

On Wed, 21 Feb 2024 at 12:37, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> Hi!
>
> > How do you suppose this would work differently from a long-lived
> > normal snapshot, which is how it works right now?
>
> Difference in the ability to take new visibility snapshot periodically
> during the second phase with rechecking visibility of tuple according
> to the "reference" snapshot (which is taken only once like now).
> It is the approach from (1) but with a workaround for the issues
> caused by heap_page_prune_opt.
>
> > Would it be exclusively for that relation?
> Yes, only for that affected relation. Other relations are unaffected.

I suppose this could work. We'd also need to be very sure that the
toast relation isn't cleaned up either: Even though that's currently
DELETE+INSERT only and can't apply HOT, it would be an issue if we
couldn't find the TOAST data of a deleted for everyone (but visible to
us) tuple.

Note that disabling cleanup for a relation will also disable cleanup
of tuple versions in that table that are not used for the R/CIC
snapshots, and that'd be an issue, too.

> > How would this be integrated with e.g. heap_page_prune_opt?
> Probably by some flag in RelationData, but not sure here yet.
>
> If the idea looks sane, I could try to extend my POC - it should be
> not too hard, likely (I already have tests to make sure it is
> correct).

I'm not a fan of this approach. Changing visibility and cleanup
semantics to only benefit R/CIC sounds like a pain to work with in
essentially all visibility-related code. I'd much rather have to deal
with another index AM, even if it takes more time: the changes in
semantics will be limited to a new plug in the index AM system and a
behaviour change in R/CIC, rather than behaviour that changes in all
visibility-checking code.

But regardless of second scan snapshots, I think we can worry about
that part at a later moment: The first scan phase is usually the most
expensive and takes the most time of all phases that hold snapshots,
and in the above discussion we agreed that we can already reduce the
time that a snapshot is held during that phase significantly. Sure, it
isn't great that we have to scan the table again with only a single
snapshot, but generally phase 2 doesn't have that much to do (except
when BRIN indexes are involved) so this is likely less of an issue.
And even if it is, we would still have reduced the number of
long-lived snapshots by half.

-Matthias

Hello!

> I'm not a fan of this approach. Changing visibility and cleanup
> semantics to only benefit R/CIC sounds like a pain to work with in
> essentially all visibility-related code. I'd much rather have to deal
> with another index AM, even if it takes more time: the changes in
> semantics will be limited to a new plug in the index AM system and a
> behaviour change in R/CIC, rather than behaviour that changes in all
> visibility-checking code.

Technically, this does not affect the visibility logic, only the
clearing semantics.
All visibility related code remains untouched.
But yes, still an inelegant and a little strange-looking option.

At the same time, perhaps it can be dressed in luxury
somehow - for example, add as a first class citizen in ComputeXidHorizonsResult
a list of blocks to clear some relations.

> But regardless of second scan snapshots, I think we can worry about
> that part at a later moment: The first scan phase is usually the most
> expensive and takes the most time of all phases that hold snapshots,
> and in the above discussion we agreed that we can already reduce the
> time that a snapshot is held during that phase significantly. Sure, it
> isn't great that we have to scan the table again with only a single
> snapshot, but generally phase 2 doesn't have that much to do (except
> when BRIN indexes are involved) so this is likely less of an issue.
> And even if it is, we would still have reduced the number of
> long-lived snapshots by half.

Hmm, but it looks like we don't have the infrastructure to "update" xmin
propagating to the horizon after the first snapshot in a transaction is taken.

One option I know of is to reuse the
d9d076222f5b94a85e0e318339cfc44b8f26022d (1) approach.
But if this is the case, then there is no point in re-taking the
snapshot again during the first
phase - just apply this "if" only for the first phase - and you're done.

Do you know any less-hacky way? Or is it a nice way to go?

[1]:
https://github.com/postgres/postgres/commit/d9d076222f5b94a85e0e318339cfc44b8f26022d#diff-8879f0173be303070ab7931db7c757c96796d84402640b9e386a4150ed97b179R1779-R1793

On Thu, 7 Mar 2024 at 19:37, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> Hello!
>
> > I'm not a fan of this approach. Changing visibility and cleanup
> > semantics to only benefit R/CIC sounds like a pain to work with in
> > essentially all visibility-related code. I'd much rather have to deal
> > with another index AM, even if it takes more time: the changes in
> > semantics will be limited to a new plug in the index AM system and a
> > behaviour change in R/CIC, rather than behaviour that changes in all
> > visibility-checking code.
>
> Technically, this does not affect the visibility logic, only the
> clearing semantics.
> All visibility related code remains untouched.

Yeah, correct. But it still needs to update the table relations'
information after finishing creating the indexes, which I'd rather not
have to do.

> But yes, still an inelegant and a little strange-looking option.
>
> At the same time, perhaps it can be dressed in luxury
> somehow - for example, add as a first class citizen in ComputeXidHorizonsResult
> a list of blocks to clear some relations.

Not sure what you mean here, but I don't think
ComputeXidHorizonsResult should have anything to do with actual
relations.

> > But regardless of second scan snapshots, I think we can worry about
> > that part at a later moment: The first scan phase is usually the most
> > expensive and takes the most time of all phases that hold snapshots,
> > and in the above discussion we agreed that we can already reduce the
> > time that a snapshot is held during that phase significantly. Sure, it
> > isn't great that we have to scan the table again with only a single
> > snapshot, but generally phase 2 doesn't have that much to do (except
> > when BRIN indexes are involved) so this is likely less of an issue.
> > And even if it is, we would still have reduced the number of
> > long-lived snapshots by half.
>
> Hmm, but it looks like we don't have the infrastructure to "update" xmin
> propagating to the horizon after the first snapshot in a transaction is taken.

We can just release the current snapshot, and get a new one, right? I
mean, we don't actually use the transaction for much else than
visibility during the first scan, and I don't think there is a need
for an actual transaction ID until we're ready to mark the index entry
with indisready.

> One option I know of is to reuse the
> d9d076222f5b94a85e0e318339cfc44b8f26022d (1) approach.
> But if this is the case, then there is no point in re-taking the
> snapshot again during the first
> phase - just apply this "if" only for the first phase - and you're done.

Not a fan of that, as it is too sensitive to abuse. Note that
extensions will also have access to these tools, and I think we should
build a system here that's not easy to break, rather than one that is.

> Do you know any less-hacky way? Or is it a nice way to go?

I suppose we could be resetting the snapshot every so often? Or use
multiple successive TID range scans with a new snapshot each?

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)

On Tue, 11 Jun 2024 at 10:58, Michail Nikolaev
<michail.nikolaev@gmail.com> wrote:
>
> Hello.
>
> I did the POC (1) of the method described in the previous email, and it looks promising.
>
> It doesn't block the VACUUM, indexes are built about 30% faster (22 mins vs 15 mins).

That's a nice improvement.

> Additional index is lightweight and does not produce any WAL.

That doesn't seem to be what I see in the current patchset:

https://github.com/postgres/postgres/compare/master...michail-nikolaev:postgres:new_index_concurrently_approach#diff-cc3cb8968cf833c4b8498ad2c561c786099c910515c4bf397ba853ae60aa2bf7R311

> I'll continue the more stress testing for a while. Also, I need to restructure the commits (my path was no direct)
intosome meaningful and reviewable patches.
 

While waiting for this, here are some initial comments on the github diffs:

- I notice you've added a new argument to
heapam_index_build_range_scan. I think this could just as well be
implemented by reading the indexInfo->ii_Concurrent field, as the
values should be equivalent, right?

- In heapam_index_build_range_scan, it seems like you're popping the
snapshot and registering a new one while holding a tuple from
heap_getnext(), thus while holding a page lock. I'm not so sure that's
OK, expecially when catalogs are also involved (specifically for
expression indexes, where functions could potentially be updated or
dropped if we re-create the visibility snapshot)

- In heapam_index_build_range_scan, you pop the snapshot before the
returned heaptuple is processed and passed to the index-provided
callback. I think that's incorrect, as it'll change the visibility of
the returned tuple before it's passed to the index's callback. I think
the snapshot manipulation is best added at the end of the loop, if we
add it at all in that function.

- The snapshot reset interval is quite high, at 500ms. Why did you
configure it that low, and didn't you make this configurable?

- You seem to be using WAL in the STIR index, while it doesn't seem
that relevant for the use case of auxiliary indexes that won't return
any data and are only used on the primary. It would imply that the
data is being sent to replicas and more data being written than
strictly necessary, which to me seems wasteful.

- The locking in stirinsert can probably be improved significantly if
we use things like atomic operations on STIR pages. We'd need an
exclusive lock only for page initialization, while share locks are
enough if the page's data is modified without WAL. That should improve
concurrent insert performance significantly, as it would further
reduce the length of the exclusively locked hot path.

Kind regards,

Matthias van de Meent
Neon (https://neon.tech)