Обсуждение: partial heap only tuples
Hello,
I'm hoping to gather some early feedback on a heap optimization I've
been working on. In short, I'm hoping to add "partial heap only
tuple" (PHOT) support, which would allow you to skip updating indexes
for unchanged columns even when other indexes require updates. Today,
HOT works wonders when no indexed columns are updated. However, as
soon as you touch one indexed column, you lose that optimization
entirely, as you must update every index on the table. The resulting
performance impact is a pain point for many of our (AWS's) enterprise
customers, so we'd like to lend a hand for some improvements in this
area. For workloads involving a lot of columns and a lot of indexes,
an optimization like PHOT can make a huge difference. I'm aware that
there was a previous attempt a few years ago to add a similar
optimization called WARM [0] [1]. However, I only noticed this
previous effort after coming up with the design for PHOT, so I ended
up taking a slightly different approach. I am also aware of a couple
of recent nbtree improvements that may mitigate some of the impact of
non-HOT updates [2] [3], but I am hoping that PHOT serves as a nice
complement to those. I've attached a very early proof-of-concept
patch with the design described below.
As far as performance is concerned, it is simple enough to show major
benefits from PHOT by tacking on a large number of indexes and columns
to a table. For a short pgbench run where each table had 5 additional
text columns and indexes on every column, I noticed a ~34% bump in
TPS with PHOT [4]. Theoretically, the TPS bump should be even higher
with additional columns with indexes. In addition to showing such
benefits, I have also attempted to show that regular pgbench runs are
not significantly affected. For a short pgbench run with no table
modifications, I noticed a ~2% bump in TPS with PHOT [5].
Next, I'll go into the design a bit. I've commandeered the two
remaining bits in t_infomask2 to use as HEAP_PHOT_UPDATED and
HEAP_PHOT_TUPLE. These are analogous to the HEAP_HOT_UPDATED and
HEAP_ONLY_TUPLE bits. (If there are concerns about exhausting the
t_infomask2 bits, I think we could only use one of the remaining bits
as a "modifier" bit on the HOT ones. I opted against that for the
proof-of-concept patch to keep things simple.) When creating a PHOT
tuple, we only create new index tuples for updated columns. These new
index tuples point to the PHOT tuple. Following is a simple
demonstration with a table with two integer columns, each with its own
index:
postgres=# SELECT heap_tuple_infomask_flags(t_infomask, t_infomask2), t_data
FROM heap_page_items(get_raw_page('test', 0))
WHERE t_infomask IS NOT NULL
OR t_infomask2 IS NOT NULL;
heap_tuple_infomask_flags | t_data
-----------------------------------------------------------------------------+--------------------
("{HEAP_XMIN_COMMITTED,HEAP_XMAX_COMMITTED,HEAP_PHOT_UPDATED}",{}) | \x0000000000000000
("{HEAP_XMIN_COMMITTED,HEAP_UPDATED,HEAP_PHOT_UPDATED,HEAP_PHOT_TUPLE}",{}) | \x0100000000000000
("{HEAP_XMAX_INVALID,HEAP_UPDATED,HEAP_PHOT_TUPLE}",{}) | \x0100000002000000
(3 rows)
postgres=# SELECT itemoffset, ctid, data
FROM bt_page_items(get_raw_page('test_a_idx', 1));
itemoffset | ctid | data
------------+-------+-------------------------
1 | (0,1) | 00 00 00 00 00 00 00 00
2 | (0,2) | 01 00 00 00 00 00 00 00
(2 rows)
postgres=# SELECT itemoffset, ctid, data
FROM bt_page_items(get_raw_page('test_b_idx', 1));
itemoffset | ctid | data
------------+-------+-------------------------
1 | (0,1) | 00 00 00 00 00 00 00 00
2 | (0,3) | 02 00 00 00 00 00 00 00
(2 rows)
When it is time to scan through a PHOT chain, there are a couple of
things to account for. Sequential scans work out-of-the-box thanks to
the visibility rules, but other types of scans like index scans
require additional checks. If you encounter a PHOT chain when
performing an index scan, you should only continue following the chain
as long as none of the columns the index indexes are modified. If the
scan does encounter such a modification, we stop following the chain
and continue with the index scan. Even if there is a tuple in that
PHOT chain that should be returned by our index scan, we will still
find it, as there will be another matching index tuple that points us
to later in the PHOT chain. My initial idea for determining which
columns were modified was to add a new bitmap after the "nulls" bitmap
in the tuple header. However, the attached patch simply uses
HeapDetermineModifiedColumns(). I've yet to measure the overhead of
this approach versus the bitmap approach, but I haven't noticed
anything too detrimental in the testing I've done so far.
In my proof-of-concept patch, I've included a temporary hack to get
some basic bitmap scans working as expected. Since we won't have
followed the PHOT chains in the bitmap index scan, we must know how to
follow them in the bitmap heap scan. Unfortunately, the bitmap heap
scan has no knowledge of what indexed columns to pay attention to when
following the PHOT chains. My temporary hack fixes this by having the
bitmap heap scan pull the set of indexed columns it needs to consider
from the outer plan. I think this is one area of the design that will
require substantially more effort. Following is a demonstration of a
basic sequential scan and bitmap scan:
postgres=# EXPLAIN (COSTS FALSE) SELECT * FROM test;
QUERY PLAN
------------------
Seq Scan on test
(1 row)
postgres=# SELECT * FROM test;
a | b
---+---
1 | 2
(1 row)
postgres=# EXPLAIN (COSTS FALSE) SELECT * FROM test WHERE a >= 0;
QUERY PLAN
---------------------------------------
Bitmap Heap Scan on test
Recheck Cond: (a >= 0)
-> Bitmap Index Scan on test_a_idx
Index Cond: (a >= 0)
(4 rows)
postgres=# SELECT * FROM test WHERE a >= 0;
a | b
---+---
1 | 2
(1 row)
This design allows for "weaving" between HOT and PHOT in a chain.
However, it is still important to treat each consecutive set of HOT
updates or PHOT updates as its own chain for the purposes of pruning
and cleanup. Pruning is heavily restricted for PHOT due to the
presence of corresponding index tuples. I believe we can redirect
line pointers for consecutive sets of PHOT updates that modify the
same set of indexed columns, but this is only possible if no index has
duplicate values in the redirected set. Also, I do not think it is
possible to prune intermediate line pointers in a PHOT chain. While
it may be possible to redirect all line pointers to the final tuple in
a series of updates to the same set of indexed columns, my hunch is
that doing so will add significant complexity for tracking
intermediate updates, and any performance gains will be marginal.
I've created some small diagrams to illustrate my proposed cleanup
strategy.
Here is a hypothetical PHOT chain.
idx1 0 1 2
idx2 0 1 2
idx3 0
lp 1 2 3 4 5
heap (0,0,0) (1,0,0) (2,0,0) (2,1,0) (2,2,0)
Heap tuples may be removed and line pointers may be redirected for
consecutive updates to the same set of indexes (as long as no index
has duplicate values in the redirected set of updates).
idx1 0 1 2
idx2 0 1 2
idx3 0
lp 1 2 -> 3 4 -> 5
heap (0,0,0) (2,0,0) (2,2,0)
When following redirect chains, we must check that the "interesting"
columns for the relevant indexes are not updated whenever a tuple is
found. In order to be eligible for cleanup, the final tuple in the
corresponding PHOT/HOT chain must also be eligible for cleanup, or all
indexes must have been updated later in the chain before any visible
tuples. (I suspect that the former condition may cause significant
bloat for some workloads and the latter condition may be prohibitively
complicated. Perhaps this can be mitigated by limiting how long we
allow PHOT chains to get.) My proof-of-concept patch does not yet
implement line pointer redirecting and cleanup, so it is possible that
I am missing some additional obstacles and optimizations here.
I think PostgreSQL 15 is realistically the earliest target version for
this change. Given that others find this project worthwhile, that's
my goal for this patch. I've CC'd a number of folks who have been
involved in this project already and who I'm hoping will continue to
help me drive this forward.
Nathan
[0] https://www.postgresql.org/message-id/flat/CABOikdMop5Rb_RnS2xFdAXMZGSqcJ-P-BY2ruMd%2BbuUkJ4iDPw%40mail.gmail.com
[1]
https://www.postgresql.org/message-id/flat/CABOikdMNy6yowA%2BwTGK9RVd8iw%2BCzqHeQSGpW7Yka_4RSZ_LOQ%40mail.gmail.com
[2] https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=0d861bbb
[3] https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=d168b666
[4] non-PHOT:
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1000
query mode: simple
number of clients: 256
number of threads: 256
duration: 1800 s
number of transactions actually processed: 29759733
latency average = 15.484 ms
latency stddev = 10.102 ms
tps = 16530.552950 (including connections establishing)
tps = 16530.730565 (excluding connections establishing)
PHOT:
...
number of transactions actually processed: 39998968
latency average = 11.520 ms
latency stddev = 8.157 ms
tps = 22220.709117 (including connections establishing)
tps = 22221.182648 (excluding connections establishing)
[5] non-PHOT:
...
number of transactions actually processed: 151841961
latency average = 3.034 ms
latency stddev = 1.854 ms
tps = 84354.268591 (including connections establishing)
tps = 84355.061353 (excluding connections establishing)
PHOT:
...
number of transactions actually processed: 155225857
latency average = 2.968 ms
latency stddev = 1.264 ms
tps = 86234.044783 (including connections establishing)
tps = 86234.961286 (excluding connections establishing)
Вложения
On Tue, Feb 9, 2021 at 06:48:21PM +0000, Bossart, Nathan wrote: > Hello, > > I'm hoping to gather some early feedback on a heap optimization I've > been working on. In short, I'm hoping to add "partial heap only > tuple" (PHOT) support, which would allow you to skip updating indexes > for unchanged columns even when other indexes require updates. Today, I think it is great you are working on this. I think it is a major way to improve performance and I have been disappointed it has not moved forward since 2016. > HOT works wonders when no indexed columns are updated. However, as > soon as you touch one indexed column, you lose that optimization > entirely, as you must update every index on the table. The resulting > performance impact is a pain point for many of our (AWS's) enterprise > customers, so we'd like to lend a hand for some improvements in this > area. For workloads involving a lot of columns and a lot of indexes, > an optimization like PHOT can make a huge difference. I'm aware that > there was a previous attempt a few years ago to add a similar > optimization called WARM [0] [1]. However, I only noticed this > previous effort after coming up with the design for PHOT, so I ended > up taking a slightly different approach. I am also aware of a couple > of recent nbtree improvements that may mitigate some of the impact of > non-HOT updates [2] [3], but I am hoping that PHOT serves as a nice > complement to those. I've attached a very early proof-of-concept > patch with the design described below. How is your approach different from those of [0] and [1]? It is interesting you still see performance benefits even after the btree duplication improvements. Did you test with those improvements? > As far as performance is concerned, it is simple enough to show major > benefits from PHOT by tacking on a large number of indexes and columns > to a table. For a short pgbench run where each table had 5 additional > text columns and indexes on every column, I noticed a ~34% bump in > TPS with PHOT [4]. Theoretically, the TPS bump should be even higher That's a big improvement. > Next, I'll go into the design a bit. I've commandeered the two > remaining bits in t_infomask2 to use as HEAP_PHOT_UPDATED and > HEAP_PHOT_TUPLE. These are analogous to the HEAP_HOT_UPDATED and > HEAP_ONLY_TUPLE bits. (If there are concerns about exhausting the > t_infomask2 bits, I think we could only use one of the remaining bits > as a "modifier" bit on the HOT ones. I opted against that for the > proof-of-concept patch to keep things simple.) When creating a PHOT > tuple, we only create new index tuples for updated columns. These new > index tuples point to the PHOT tuple. Following is a simple > demonstration with a table with two integer columns, each with its own > index: Whatever solution you have, you have to be able to handle adding/removing columns, and adding/removing indexes. > When it is time to scan through a PHOT chain, there are a couple of > things to account for. Sequential scans work out-of-the-box thanks to > the visibility rules, but other types of scans like index scans > require additional checks. If you encounter a PHOT chain when > performing an index scan, you should only continue following the chain > as long as none of the columns the index indexes are modified. If the > scan does encounter such a modification, we stop following the chain > and continue with the index scan. Even if there is a tuple in that I think in patch [0] and [1], if an index column changes, all the indexes had to be inserted into, while you seem to require inserts only into the index that needs it. Is that correct? > PHOT chain that should be returned by our index scan, we will still > find it, as there will be another matching index tuple that points us > to later in the PHOT chain. My initial idea for determining which > columns were modified was to add a new bitmap after the "nulls" bitmap > in the tuple header. However, the attached patch simply uses > HeapDetermineModifiedColumns(). I've yet to measure the overhead of > this approach versus the bitmap approach, but I haven't noticed > anything too detrimental in the testing I've done so far. A bitmap is an interesting approach, but you are right it will need benchmarking. I wonder if you should create a Postgres wiki page to document all of this. I agree PG 15 makes sense. I would like to help with this if I can. I will need to study this email more later. -- Bruce Momjian <bruce@momjian.us> https://momjian.us EDB https://enterprisedb.com The usefulness of a cup is in its emptiness, Bruce Lee
On 2/10/21, 2:43 PM, "Bruce Momjian" <bruce@momjian.us> wrote: > On Tue, Feb 9, 2021 at 06:48:21PM +0000, Bossart, Nathan wrote: >> HOT works wonders when no indexed columns are updated. However, as >> soon as you touch one indexed column, you lose that optimization >> entirely, as you must update every index on the table. The resulting >> performance impact is a pain point for many of our (AWS's) enterprise >> customers, so we'd like to lend a hand for some improvements in this >> area. For workloads involving a lot of columns and a lot of indexes, >> an optimization like PHOT can make a huge difference. I'm aware that >> there was a previous attempt a few years ago to add a similar >> optimization called WARM [0] [1]. However, I only noticed this >> previous effort after coming up with the design for PHOT, so I ended >> up taking a slightly different approach. I am also aware of a couple >> of recent nbtree improvements that may mitigate some of the impact of >> non-HOT updates [2] [3], but I am hoping that PHOT serves as a nice >> complement to those. I've attached a very early proof-of-concept >> patch with the design described below. > > How is your approach different from those of [0] and [1]? It is > interesting you still see performance benefits even after the btree > duplication improvements. Did you test with those improvements? I believe one of the main differences is that index tuples will point to the corresponding PHOT tuple instead of the root of the HOT/PHOT chain. I'm sure there are other differences. I plan on giving those two long threads another read-through in the near future. I made sure that the btree duplication improvements were applied for my benchmarking. IIUC those don't alleviate the requirement that you insert all index tuples for non-HOT updates, so PHOT can still provide some added benefits there. >> Next, I'll go into the design a bit. I've commandeered the two >> remaining bits in t_infomask2 to use as HEAP_PHOT_UPDATED and >> HEAP_PHOT_TUPLE. These are analogous to the HEAP_HOT_UPDATED and >> HEAP_ONLY_TUPLE bits. (If there are concerns about exhausting the >> t_infomask2 bits, I think we could only use one of the remaining bits >> as a "modifier" bit on the HOT ones. I opted against that for the >> proof-of-concept patch to keep things simple.) When creating a PHOT >> tuple, we only create new index tuples for updated columns. These new >> index tuples point to the PHOT tuple. Following is a simple >> demonstration with a table with two integer columns, each with its own >> index: > > Whatever solution you have, you have to be able to handle > adding/removing columns, and adding/removing indexes. I admittedly have not thought too much about the implications of adding/removing columns and indexes for PHOT yet, but that's definitely an important part of this project that I need to look into. I see that HOT has some special handling for commands like CREATE INDEX that I can reference. >> When it is time to scan through a PHOT chain, there are a couple of >> things to account for. Sequential scans work out-of-the-box thanks to >> the visibility rules, but other types of scans like index scans >> require additional checks. If you encounter a PHOT chain when >> performing an index scan, you should only continue following the chain >> as long as none of the columns the index indexes are modified. If the >> scan does encounter such a modification, we stop following the chain >> and continue with the index scan. Even if there is a tuple in that > > I think in patch [0] and [1], if an index column changes, all the > indexes had to be inserted into, while you seem to require inserts only > into the index that needs it. Is that correct? Right, PHOT only requires new index tuples for the modified columns. However, I was under the impression that WARM aimed to do the same thing. I might be misunderstanding your question. > I wonder if you should create a Postgres wiki page to document all of > this. I agree PG 15 makes sense. I would like to help with this if I > can. I will need to study this email more later. Thanks for taking a look. I think a wiki is a good idea for keeping track of the current state of the design. I'll look into that. Nathan
Hi, On 2021-02-09 18:48:21 +0000, Bossart, Nathan wrote: > In order to be eligible for cleanup, the final tuple in the > corresponding PHOT/HOT chain must also be eligible for cleanup, or all > indexes must have been updated later in the chain before any visible > tuples. This sounds like it might be prohibitively painful. Adding effectively unremovable bloat to remove other bloat is not an uncomplicated premise. I think you'd really need a way to fully remove this as part of vacuum for this to be viable. Greetings, Andres Freund
On 2/13/21, 8:26 AM, "Andres Freund" <andres@anarazel.de> wrote:
> On 2021-02-09 18:48:21 +0000, Bossart, Nathan wrote:
>> In order to be eligible for cleanup, the final tuple in the
>> corresponding PHOT/HOT chain must also be eligible for cleanup, or all
>> indexes must have been updated later in the chain before any visible
>> tuples.
>
> This sounds like it might be prohibitively painful. Adding effectively
> unremovable bloat to remove other bloat is not an uncomplicated
> premise. I think you'd really need a way to fully remove this as part of
> vacuum for this to be viable.
Yeah, this is something I'm concerned about. I think adding a bitmap
of modified columns to the header of PHOT-updated tuples improves
matters quite a bit, even for single-page vacuuming. Following is a
strategy I've been developing (there may still be some gaps). Here's
a basic PHOT chain where all tuples are visible and the last one has
not been deleted or updated:
idx1 0 1 2 3
idx2 0 1 2
idx3 0 2 3
lp 1 2 3 4 5
tuple (0,0,0) (0,1,1) (2,2,1) (2,2,2) (3,2,3)
bitmap -xx xx- --x x-x
For single-page vacuum, we take the following actions:
1. Starting at the root of the PHOT chain, create an OR'd bitmap
of the chain.
2. Walk backwards, OR-ing the bitmaps. Stop when the bitmap
matches the one from step 1. As we walk backwards, identify
"key" tuples, which are tuples where the OR'd bitmap changes as
you walk backwards. If the OR'd bitmap does not include all
columns for the table, also include the root of the PHOT chain
as a key tuple.
3. Redirect each key tuple to the next key tuple.
4. For all but the first key tuple, OR the bitmaps of all pruned
tuples from each key tuple (exclusive) to the next key tuple
(inclusive) and store the result in the bitmap of the next key
tuple.
5. Mark all line pointers for all non-key tuples as dead. Storage
can be removed for all tuples except the last one, but we must
leave around the bitmap for all key tuples except for the first
one.
After this, our basic PHOT chain looks like this:
idx1 0 1 2 3
idx2 0 1 2
idx3 0 2 3
lp X X 3->5 X 5
tuple (3,2,3)
bitmap x-x
Without PHOT, this intermediate state would have 15 index tuples, 5
line pointers, and 1 heap tuples. With PHOT, we have 10 index tuples,
5 line pointers, 1 heap tuple, and 1 bitmap. When we vacuum the
indexes, we can reclaim the dead line pointers and remove the
associated index tuples:
idx1 3
idx2 2
idx3 2 3
lp 3->5 5
tuple (3,2,3)
bitmap x-x
Without PHOT, this final state would have 3 index tuples, 1 line
pointer, and 1 heap tuple. With PHOT, we have 4 index tuples, 2 line
pointers, 1 heap tuple, and 1 bitmap. Overall, we still end up
keeping around more line pointers and tuple headers (for the bitmaps),
but maybe that is good enough. I think the next step here would be to
find a way to remove some of the unnecessary index tuples and adjust
the remaining ones to point to the last line pointer in the PHOT
chain.
Nathan
On Tue, Feb 9, 2021 at 10:48 AM Bossart, Nathan <bossartn@amazon.com> wrote: > I'm hoping to gather some early feedback on a heap optimization I've > been working on. In short, I'm hoping to add "partial heap only > tuple" (PHOT) support, which would allow you to skip updating indexes > for unchanged columns even when other indexes require updates. Today, > HOT works wonders when no indexed columns are updated. However, as > soon as you touch one indexed column, you lose that optimization > entirely, as you must update every index on the table. The resulting > performance impact is a pain point for many of our (AWS's) enterprise > customers, so we'd like to lend a hand for some improvements in this > area. For workloads involving a lot of columns and a lot of indexes, > an optimization like PHOT can make a huge difference. I'm aware that > there was a previous attempt a few years ago to add a similar > optimization called WARM [0] [1]. However, I only noticed this > previous effort after coming up with the design for PHOT, so I ended > up taking a slightly different approach. I am also aware of a couple > of recent nbtree improvements that may mitigate some of the impact of > non-HOT updates [2] [3], but I am hoping that PHOT serves as a nice > complement to those. I've attached a very early proof-of-concept > patch with the design described below. I would like to share some thoughts that I have about how I think about optimizations like PHOT, and how they might fit together with my own work -- particularly the nbtree bottom-up index deletion feature you referenced. My remarks could equally well apply to WARM. Ordinarily this is the kind of thing that would be too hand-wavey for the mailing list, but we don't have the luxury of in-person communication right now. Everybody tends to talk about HOT as if it works perfectly once you make some modest assumptions, such as "there are no long-running transactions", and "no UPDATEs will logically modify indexed columns". But I tend to doubt that that's truly the case -- I think that there are still pathological cases where HOT cannot keep the total table size stable in the long run due to subtle effects that eventually aggregate into significant issues, like heap fragmentation. Ask Jan Wieck about the stability of some of the TPC-C/BenchmarkSQL tables to get one example of this. There is no reason to believe that PHOT will help with that. Maybe that's okay, but I would think carefully about what that means if I were undertaking this work. Ensuring stability in the on-disk size of tables in cases where the size of the logical database is stable should be an important goal of a project like PHOT or HOT. If you want to get a better sense of how these inefficiencies might happen, I suggest looking into using recently added autovacuum logging to analyze how well HOT works today, using the technique that I go into here: https://postgr.es/m/CAH2-WzkjU+NiBskZunBDpz6trSe+aQvuPAe+xgM8ZvoB4wQwhA@mail.gmail.com Small inefficiencies in the on-disk structure have a tendency to aggregate over time, at least when there is no possible way to reverse them. The bottom-up index deletion stuff is very effective as a backstop against index bloat, because things are generally very non-linear. The cost of an unnecessary page split is very high, and permanent. But we can make it cheap to *try* to avoid that using fairly simple heuristics. We can be reasonably confident that we're about to split the page unnecessarily, and use cues that ramp up the number of heap page accesses as needed. We ramp up during a bottom-up index deletion, as we manage to free some index tuples as a result of previous heap page accesses. This works very well because we can intervene very selectively. We aren't interested in deleting index tuples unless and until we really have to, and in general there tends to be quite a bit of free space to temporarily store extra version duplicates -- that's what most index pages look like, even on the busiest of databases. It's possible for the bottom-up index deletion mechanism to be invoked very infrequently, and yet make a huge difference. And when it fails to free anything, it fails permanently for that particular leaf page (because it splits) -- so now we have lots of space for future index tuple insertions that cover the original page's key space. We won't thrash. My intuition is that similar principles can be applied inside heapam. Failing to fit related versions on a heap page (having managed to do so for hours or days before that point) is more or less the heap page equivalent of a leaf page split from version churn (this is the pathology that bottom-up index deletion targets). For example, we could have a fall back mode that compresses old versions that is used if and only if heap pruning was attempted but then failed. We should always try to avoid migrating to a new heap page, because that amounts to a permanent solution to a temporary problem. We should perhaps make the updater work to prove that that's truly necessary, rather than giving up immediately (i.e. assuming that it must be necessary at the first sign of trouble). We might have successfully fit the successor heap tuple version a million times before just by HOT pruning, and yet currently we give up just because it didn't work on the one millionth and first occasion -- don't you think that's kind of silly? We may be able to afford having a fallback strategy that is relatively expensive, provided it is rarely used. And it might be very effective in the aggregate, despite being rarely used -- it might provide us just what we were missing before. Just try harder when you run into a problem every once in a blue moon! A diversity of strategies with fallback behavior is sometimes the best strategy. Don't underestimate the contribution of rare and seemingly insignificant adverse events. Consider the lifecycle of the data over time. If we quit trying to fit new versions on the same heap page at the first sign of real trouble, then it's only a matter of time until widespread heap fragmentation results -- each heap page only has to be unlucky once, and in the long run it's inevitable that they all will. We could probably do better at nipping it in the bud at the level of individual heap pages and opportunistic prune operations. I'm sure that it would be useful to not have to rely on bottom-up index deletion in more cases -- I think that the idea of "a better HOT" might still be very helpful. Bottom-up index deletion is only supposed to be a backstop against pathological behavior (version churn page splits), which is probably always going to be possible with a sufficiently extreme workload. I don't believe that the current levels of version churn/write amplification that we still see with Postgres must be addressed through totally eliminating multiple versions of the same logical row that live together in the same heap page. This idea is a false dichotomy. And it fails to acknowledge how the current design often works very well. When and how it fails to work well with a real workload and real tuning (especially heap fill factor tuning) is probably not well understood. Why not start with that? Our default heap fill factor is 100. Maybe that's the right decision, but it significantly impedes the ability of HOT to keep the size of tables stable over time. Just because heap fill factor 90 also has issues today doesn't mean that each pathological behavior cannot be fixed through targeted intervention. Maybe the myth that HOT works perfectly once you make some modest assumptions could come true. -- Peter Geoghegan
On Sun, Apr 18, 2021 at 04:27:15PM -0700, Peter Geoghegan wrote: > Everybody tends to talk about HOT as if it works perfectly once you > make some modest assumptions, such as "there are no long-running > transactions", and "no UPDATEs will logically modify indexed columns". > But I tend to doubt that that's truly the case -- I think that there > are still pathological cases where HOT cannot keep the total table > size stable in the long run due to subtle effects that eventually > aggregate into significant issues, like heap fragmentation. Ask Jan > Wieck about the stability of some of the TPC-C/BenchmarkSQL tables to ... > We might have successfully fit the successor heap tuple version a > million times before just by HOT pruning, and yet currently we give up > just because it didn't work on the one millionth and first occasion -- > don't you think that's kind of silly? We may be able to afford having > a fallback strategy that is relatively expensive, provided it is > rarely used. And it might be very effective in the aggregate, despite > being rarely used -- it might provide us just what we were missing > before. Just try harder when you run into a problem every once in a > blue moon! > > A diversity of strategies with fallback behavior is sometimes the best > strategy. Don't underestimate the contribution of rare and seemingly > insignificant adverse events. Consider the lifecycle of the data over That is an intersting point --- we often focus on optimizing frequent operations, but preventing rare but expensive-in-aggregate events from happening is also useful. -- Bruce Momjian <bruce@momjian.us> https://momjian.us EDB https://enterprisedb.com If only the physical world exists, free will is an illusion.
On Mon, Apr 19, 2021 at 5:09 PM Bruce Momjian <bruce@momjian.us> wrote: > > A diversity of strategies with fallback behavior is sometimes the best > > strategy. Don't underestimate the contribution of rare and seemingly > > insignificant adverse events. Consider the lifecycle of the data over > > That is an intersting point --- we often focus on optimizing frequent > operations, but preventing rare but expensive-in-aggregate events from > happening is also useful. Right. Similarly, we sometimes focus on adding an improvement, overlooking more promising opportunities to subtract a disimprovement. Apparently this is a well known tendency: https://www.scientificamerican.com/article/our-brain-typically-overlooks-this-brilliant-problem-solving-strategy/ I believe that it's particularly important to consider subtractive approaches with a complex system. This has sometimes worked well for me as a conscious and deliberate strategy. -- Peter Geoghegan
On Tue, Mar 9, 2021 at 12:09 AM Bossart, Nathan <bossartn@amazon.com> wrote: > > On 3/8/21, 10:16 AM, "Ibrar Ahmed" <ibrar.ahmad@gmail.com> wrote: > > On Wed, Feb 24, 2021 at 3:22 AM Bossart, Nathan <bossartn@amazon.com> wrote: > >> On 2/10/21, 2:43 PM, "Bruce Momjian" <bruce@momjian.us> wrote: > >>> I wonder if you should create a Postgres wiki page to document all of > >>> this. I agree PG 15 makes sense. I would like to help with this if I > >>> can. I will need to study this email more later. > >> > >> I've started the wiki page for this: > >> > >> https://wiki.postgresql.org/wiki/Partial_Heap_Only_Tuples > >> > >> Nathan > > > > The regression test case (partial-index) is failing > > > > https://cirrus-ci.com/task/5310522716323840 > > This patch is intended as a proof-of-concept of some basic pieces of > the project. I'm working on a new patch set that should be more > suitable for community review. The patch does not apply on Head anymore, could you rebase and post a patch. I'm changing the status to "Waiting for Author". Regards, Vignesh
> On 14 Jul 2021, at 13:34, vignesh C <vignesh21@gmail.com> wrote: > The patch does not apply on Head anymore, could you rebase and post a > patch. I'm changing the status to "Waiting for Author". As no update has been posted, the patch still doesn't apply. I'm marking this patch Returned with Feedback, feel free to open a new entry for an updated patch. -- Daniel Gustafsson https://vmware.com/
On 11/4/21, 3:24 AM, "Daniel Gustafsson" <daniel@yesql.se> wrote: > As no update has been posted, the patch still doesn't apply. I'm marking this > patch Returned with Feedback, feel free to open a new entry for an updated > patch. Thanks. I have been working on this intermittently, and I hope to post a more complete proof-of-concept in the near future. I'll create a new commitfest entry once that's done. Nathan