Previously, the decision to store values in leaves or within the child
pointer was made at compile time, with variable length values using
leaves by necessity. This commit allows introspecting the length of
variable length values at runtime for that decision. This requires
the ability to tell whether the last-level child pointer is actually
a value, so we use a pointer tag in the lowest level bit.
Use this in TID store. This entails adding a byte to the header to
reserve space for the tag. Commit f35bd9bf3 stores up to three offsets
within the header with no bitmap, and now the header can be embedded
as above. This reduces worst-case memory usage when TIDs are sparse.
Reviewed (in an earlier version) by Masahiko Sawada
Discussion: https://postgr.es/m/CANWCAZYw+_KAaUNruhJfE=h6WgtBKeDG32St8vBJBEY82bGVRQ@mail.gmail.com
Discussion: https://postgr.es/m/CAD21AoBci3Hujzijubomo1tdwH3XtQ9F89cTNQ4bsQijOmqnEw@mail.gmail.com
While SH_STAT() is only used for debugging, the allocated array can be large,
and therefore should be freed.
It's unclear why coverity started warning now.
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Reported-by: Coverity
Discussion: https://postgr.es/m/3005248.1712538233@sss.pgh.pa.us
Backpatch: 12-
06c418e163 introduced pg_wal_replay_wait() procedure allowing to wait for
the particular LSN to be replayed on standby. The waiters were stored in
the flat array. Even though scanning small arrays is fast, that might be a
problem at scale (a lot of waiting processes).
This commit replaces the flat shared memory array with the pairing heap,
which holds the waiter with the least LSN at the top. This gives us O(log N)
complexity for both inserting and removing waiters.
Reported-by: Alvaro Herrera
Discussion: https://postgr.es/m/202404030658.hhj3vfxeyhft%40alvherre.pgsql
Previously, binaryheap didn't support updating a key and removing a
node in an efficient way. For example, in order to remove a node from
the binaryheap, the caller had to pass the node's position within the
array that the binaryheap internally has. Removing a node from the
binaryheap is done in O(log n) but searching for the key's position is
done in O(n).
This commit adds a hash table to binaryheap in order to track the
position of each nodes in the binaryheap. That way, by using newly
added functions such as binaryheap_update_up() etc., both updating a
key and removing a node can be done in O(1) on an average and O(log n)
in worst case. This is known as the indexed binary heap. The caller
can specify to use the indexed binaryheap by passing indexed = true.
The current code does not use the new indexing logic, but it will be
used by an upcoming patch.
Reviewed-by: Vignesh C, Peter Smith, Hayato Kuroda, Ajin Cherian,
Tomas Vondra, Shubham Khanna
Discussion: https://postgr.es/m/CAD21AoDffo37RC-eUuyHJKVEr017V2YYDLyn1xF_00ofptWbkg%40mail.gmail.com
Coverity complained about the integer handling issue; if we start with
an arbitrary non-negative shift value, the loop may decrement it down
to something less than zero before exiting. This commit adds an
assertion to make sure the 'shift' is always 0 after the loop, and
uses 0 as the shift to get the key chunk in the following operation.
Introduced by ee1b30f12.
Reported-by: Tom Lane as per coverity
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/2089517.1711299216%40sss.pgh.pa.us
destroyStringInfo() is a counterpart to makeStringInfo(), freeing a
palloc'd StringInfo and its data. This is a convenience function to
align the StringInfo API with the PQExpBuffer API. Originally added
in the OAuth patchset, it was extracted and committed separately in
order to aid upcoming JSON work.
Author: Daniel Gustafsson <daniel@yesql.se>
Author: Jacob Champion <jacob.champion@enterprisedb.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAOYmi+mWdTd6ujtyF7MsvXvk7ToLRVG_tYAcaGbQLvf=N4KrQw@mail.gmail.com
This implements a radix tree data structure based on the design in
"The Adaptive Radix Tree: ARTful Indexing for Main-Memory Databases"
by Viktor Leis, Alfons Kemper, and ThomasNeumann, 2013. The main
technique that makes it adaptive is using several different node types,
each with a different capacity of elements, and a different algorithm
for accessing them. The nodes start small and grow/shrink as needed.
The main advantage over hash tables is efficient sorted iteration and
better memory locality when successive keys are lexicographically
close together. The implementation currently assumes 64-bit integer
keys, and traversing the tree is in general slower than a linear
probing hash table, so this is not a general-purpose associative array.
The paper describes two other techniques not implemented here,
namely "path compression" and "lazy expansion". These can further
reduce memory usage and speed up traversal, but the former would add
significant complexity and the latter requires storing the full key
with the value. We do trivially compress the path when leading bytes
of the key are zeros, however.
For value storage, we use "combined pointer/value slots", as
recommended in the paper. Values of size equal or smaller than the the
platform's pointer type are stored in the array of child pointers in
the last level node, while larger values are each stored in a separate
allocation. This is for now fixed at compile time, but it would be
fairly trivial to allow determining at runtime how variable-length
values are stored.
One innovation in our implementation compared to the ART paper is
decoupling the notion of node "size class" from "kind". The size
classes within a given node kind have the same underlying type, but
a variable capacity for children, so we can introduce additional node
sizes with little additional code.
To enable different use cases to specialize for different value types
and for shared/local memory, we use macro-templatized code generation
in the same manner as simplehash.h and sort_template.h.
Future commits will use this infrastructure for storing TIDs.
Patch by Masahiko Sawada and John Naylor, but a substantial amount of
credit is due to Andres Freund, whose proof-of-concept was a valuable
source of coding idioms and awareness of performance pitfalls, and
who reviewed earlier versions.
Discussion: https://postgr.es/m/CAD21AoAfOZvmfR0j8VmZorZjL7RhTiQdVttNuC4W-Shdc2a-AA%40mail.gmail.com
A recent commit added a copy_function member to the
dshash_parameters struct, but it missed updating a couple of
comments that refer to the function pointer members of this struct.
One of those comments also refers to a tranche_name member and non-
arg variants of the function pointer members, all of which were
either removed during development or removed shortly after dshash
table support was committed.
Oversights in commits 8c0d7bafad, d7694fc148, and 42a1de3013.
Discussion: https://postgr.es/m/20240227045213.GA2329190%40nathanxps13
Presently, string keys are not well-supported for dshash tables.
The dshash code always copies key_size bytes into new entries'
keys, and dshash.h only provides compare and hash functions that
forward to memcmp() and tag_hash(), both of which do not stop at
the first NUL. This means that callers must pad string keys so
that the data beyond the first NUL does not adversely affect the
results of copying, comparing, and hashing the keys.
To better support string keys in dshash tables, this commit does
a couple things:
* A new copy_function field is added to the dshash_parameters
struct. This function pointer specifies how the key should be
copied into new table entries. For example, we only want to copy
up to the first NUL byte for string keys. A dshash_memcpy()
helper function is provided and used for all existing in-tree
dshash tables without string keys.
* A set of helper functions for string keys are provided. These
helper functions forward to strcmp(), strcpy(), and
string_hash(), all of which ignore data beyond the first NUL.
This commit also adjusts the DSM registry's dshash table to use the
new helper functions for string keys.
Reviewed-by: Andy Fan
Discussion: https://postgr.es/m/20240119215941.GA1322079%40nathanxps13
This commit adds integer comparison functions that are designed to
be as efficient as possible while avoiding overflow. A follow-up
commit will make use of these functions in many of the in-tree
qsort() comparators. The new functions are not better in all cases
(e.g., when the comparator function is inlined), so it is important
to consider the context before using them.
Author: Mats Kindahl
Reviewed-by: Tom Lane, Heikki Linnakangas, Andres Freund, Thomas Munro, Andrey Borodin, Fabrízio de Royes Mello
Discussion: https://postgr.es/m/CA%2B14426g2Wa9QuUpmakwPxXFWG_1FaY0AsApkvcTBy-YfS6uaw%40mail.gmail.com
Compute size first, then allocate, then update the structure.
Previously, an out-of-memory when growing could leave the hashtable in
an inconsistent state.
Discussion: https://postgr.es/m/20231117201334.eyb542qr5yk4gilq@awork3.anarazel.de
Reviewed-by: Andres Freund
Reviewed-by: Gurjeet Singh
There were various places in our codebase which conjured up a StringInfo
by manually assigning the StringInfo fields and setting the data field
to point to some existing buffer. There wasn't much consistency here as
to what fields like maxlen got set to and in one location we didn't
correctly ensure that the buffer was correctly NUL terminated at len
bytes, as per what was documented as required in stringinfo.h
Here we introduce 2 new functions to initialize StringInfos. One allows
callers to initialize a StringInfo passing along a buffer that is
already allocated by palloc. Here the StringInfo code uses this buffer
directly rather than doing any memcpying into a new allocation. Having
this as a function allows us to verify the buffer is correctly NUL
terminated. StringInfos initialized this way can be appended to and
reset just like any other normal StringInfo.
The other new initialization function also accepts an existing buffer,
but the given buffer does not need to be a pointer to a palloc'd chunk.
This buffer could be a pointer pointing partway into some palloc'd chunk
or may not even be palloc'd at all. StringInfos initialized this way
are deemed as "read-only". This means that it's not possible to
append to them or reset them.
For the latter of the two new initialization functions mentioned above,
we relax the requirement that the data buffer must be NUL terminated.
Relaxing this requirement is convenient in a few places as it can save
us from having to allocate an entire new buffer just to add the NUL
terminator or save us from having to temporarily add a NUL only to have to
put the original char back again later.
Incompatibility note:
Here we also forego adding the NUL in a few places where it does not
seem to be required. These locations are passing the given StringInfo
into a type's receive function. It does not seem like any of our
built-in receive functions require this, but perhaps there's some UDT
out there in the wild which does require this. It is likely worthy of
a mention in the release notes that a UDT's receive function mustn't rely
on the input StringInfo being NUL terminated.
Author: David Rowley
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CAApHDvorfO3iBZ%3DxpiZvp3uHtJVLyFaPBSvcAhAq2HPLnaNSwQ%40mail.gmail.com
This commit introduces binaryheap_remove_node(), which can be used
to remove any node from a binary heap. The implementation is
straightforward. The target node is replaced with the last node in
the heap, and then we sift as needed to preserve the heap property.
This new function is intended for use in a follow-up commit that
will improve the performance of pg_restore.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/3612876.1689443232%40sss.pgh.pa.us
There are a couple of places in frontend code that could make use
of this simple binary heap implementation. This commit makes
binaryheap usable in frontend code, much like commit 26aaf97b68 did
for StringInfo. Like StringInfo, the header file is left in lib/
to reduce the likelihood of unnecessary breakage.
The frontend version of binaryheap exposes a void *-based API since
frontend code does not have access to the Datum definitions. This
seemed like a better approach than switching all existing uses to
void * or making the Datum definitions available to frontend code.
Reviewed-by: Tom Lane, Alvaro Herrera
Discussion: https://postgr.es/m/3612876.1689443232%40sss.pgh.pa.us
Invent separate macros for "invalid" values of these types, so that
we needn't embed knowledge of their representations into calling code.
These are all zeroes anyway ATM, so this is not fixing any live bug,
but it makes the code cleaner and more future-proof.
I (tgl) also chose to move DSM_HANDLE_INVALID into dsm_impl.h,
since it seems like it should live beside the typedef for dsm_handle.
Hou Zhijie, Nathan Bossart, Kyotaro Horiguchi, Tom Lane
Discussion: https://postgr.es/m/OS0PR01MB5716860B1454C34E5B179B6694C99@OS0PR01MB5716.jpnprd01.prod.outlook.com
Const qualifiers ensure that we don't do something stupid in the
function implementation. Additionally they clarify the interface. As
an example:
void
slist_delete(slist_head *head, const slist_node *node)
Here one can instantly tell that node->next is not going to be set to
NULL. Finally, const qualifiers potentially allow the compiler to do
more optimizations. This being said, no benchmarking was done for
this patch.
The functions that return non-const pointers like slist_next_node(),
dclist_next_node() etc. are not affected by the patch intentionally.
Author: Aleksander Alekseev
Reviewed-by: Andres Freund
Discussion: https://postgr.es/m/CAJ7c6TM2%3D08mNKD9aJg8vEY9hd%2BG4L7%2BNvh30UiNT3kShgRgNg%40mail.gmail.com
While on it, newlines are removed from the end of two elog() strings.
The others are simple grammar mistakes. One comment in pg_upgrade
referred incorrectly to sequences since a7e5457.
Author: Justin Pryzby
Discussion: https://postgr.es/m/20221230231257.GI1153@telsasoft.com
Backpatch-through: 11
Casting the result of palloc etc. to the intended type is more per
project style anyway.
(The fact that cpluspluscheck doesn't notice these problems is
because it doesn't expand any macros, which seems like a troubling
shortcoming. Don't have a good idea about improving that.)
Back-patch to v13, which is as far as the patch applies cleanly;
doesn't seem worth working harder.
David Geier
Discussion: https://postgr.es/m/aa5d88a3-71f4-3455-11cf-82de0372c941@gmail.com
We have various requirements when using a dlist_head to keep track of the
number of items in the list. This, traditionally, has been done by
maintaining a counter variable in the calling code. Here we tidy this up
by adding "dclist", which is very similar to dlist but also keeps track of
the number of items stored in the list.
Callers may use the new dclist_count() function when they need to know how
many items are stored. Obtaining the count is an O(1) operation.
For simplicity reasons, dclist and dlist both use dlist_node as their node
type and dlist_iter/dlist_mutable_iter as their iterator type. dclists
have all of the same functionality as dlists except there is no function
named dclist_delete(). To remove an item from a list dclist_delete_from()
must be used. This requires knowing which dclist the given item is stored
in.
Additionally, here we also convert some dlists where additional code
exists to keep track of the number of items stored and to make these use
dclists instead.
Author: David Rowley
Reviewed-by: Bharath Rupireddy, Aleksander Alekseev
Discussion: https://postgr.es/m/CAApHDvrtVxr+FXEX0VbViCFKDGxA3tWDgw9oFewNXCJMmwLjLg@mail.gmail.com
In a similar effort to f01592f91, here we mostly rename shadowed local
variables to remove the warnings produced when compiling with
-Wshadow=compatible-local.
This fixes 63 warnings and leaves just 5.
Author: Justin Pryzby, David Rowley
Reviewed-by: Justin Pryzby
Discussion https://postgr.es/m/20220817145434.GC26426%40telsasoft.com
PostgreSQL contains the implementation of the red-black tree. The red-black
tree is the ordered data structure, and one of its advantages is the ability
to do inequality searches. This commit adds rbt_find_less() and
rbt_find_great() functions implementing these searches. While these searches
aren't yet used in the core code, they might be useful for extensions.
Discussion: https://postgr.es/m/CAGRrpzYE8-7GCoaPjOiL9T_HY605MRax-2jgTtLq236uksZ1Sw%40mail.gmail.com
Author: Steve Chavez, Alexander Korotkov
Reviewed-by: Alexander Korotkov
Get rid of the separate "FATAL" log level, as it was applied
so inconsistently as to be meaningless. This mostly involves
s/pg_log_fatal/pg_log_error/g.
Create a macro pg_fatal() to handle the common use-case of
pg_log_error() immediately followed by exit(1). Various
modules had already invented either this or equivalent macros;
standardize on pg_fatal() and apply it where possible.
Invent the ability to add "detail" and "hint" messages to a
frontend message, much as we have long had in the backend.
Except where rewording was needed to convert existing coding
to detail/hint style, I have (mostly) resisted the temptation
to change existing message wording.
Patch by me. Design and patch reviewed at various stages by
Robert Haas, Kyotaro Horiguchi, Peter Eisentraut and
Daniel Gustafsson.
Discussion: https://postgr.es/m/1363732.1636496441@sss.pgh.pa.us
The previous coding of dshash_seq_next(), on the first call, accessed
status->hash_table->size_log2 without holding a partition lock and without
guaranteeing that ensure_valid_bucket_pointers() had ever been called.
That oversight turns out to not have immediately visible effects, because
bucket 0 is always in partition 0, and ensure_valid_bucket_pointers() was
called after acquiring the partition lock. However,
PARTITION_FOR_BUCKET_INDEX() with a size_log2 of 0 ends up triggering formally
undefined behaviour.
Simplify by accessing partition 0, without using PARTITION_FOR_BUCKET_INDEX().
While at it, remove dshash_get_current(), there is no convincing use
case. Also polish a few comments.
Author: Andres Freund <andres@anarazel.de>
Reviewed-By: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/CA+hUKGL9hY_VY=+oUK+Gc1iSRx-Ls5qeYJ6q=dQVZnT3R63Taw@mail.gmail.com
Add ability to scan all entries sequentially to dshash. The interface is
similar but a bit different both from that of dynahash and simple dshash
search functions. The most significant differences is that dshash's interfac
always needs a call to dshash_seq_term when scan ends. Another is
locking. Dshash holds partition lock when returning an entry,
dshash_seq_next() also holds lock when returning an entry but callers
shouldn't release it, since the lock is essential to continue a scan. The
seqscan interface allows entry deletion while a scan is in progress using
dshash_delete_current().
Reviewed-By: Andres Freund <andres@anarazel.de>
Author: Kyotaro Horiguchi <horikyoga.ntt@gmail.com>
The code does not expect sh_error() to return, but the patch
that made this header usable in frontend didn't get that memo.
While here, plaster unlikely() on the tests that decide whether
to invoke sh_error(), and add our standard copyright notice.
Noted by Andres Freund. Back-patch to v13 where this frontend
support came in.
Discussion: https://postgr.es/m/0D54435C-1199-4361-9D74-2FBDCF8EA164@anarazel.de
I just got burnt by trying to use pg_malloc instead of pg_malloc0
with this. Save the next hacker some time by not leaving this
API detail undocumented.
This fixes a bug in simplehash.h which caused an incorrect size mask to be
used when the hash table grew to SH_MAX_SIZE (2^32). The code was
incorrectly setting the size mask to 0 when the hash tables reached the
maximum possible number of buckets. This would result always trying to
use the 0th bucket causing an infinite loop of trying to grow the hash
table due to there being too many collisions.
Seemingly it's not that common for simplehash tables to ever grow this big
as this bug dates back to v10 and nobody seems to have noticed it before.
However, probably the most likely place that people would notice it would
be doing a large in-memory Hash Aggregate with something close to at least
2^31 groups.
After this fix, the code now works correctly with up to within 98% of 2^32
groups and will fail with the following error when trying to insert any
more items into the hash table:
ERROR: hash table size exceeded
However, the work_mem (or hash_mem_multiplier in newer versions) settings
will generally cause Hash Aggregates to spill to disk long before reaching
that many groups. The minimal test case I did took a work_mem setting of
over 192GB to hit the bug.
simplehash hash tables are used in a few other places such as Bitmap Index
Scans, however, again the size that the hash table can become there is
also limited to work_mem and it would take a relation of around 16TB
(2^31) pages and a very large work_mem setting to hit this. With smaller
work_mem values the table would become lossy and never grow large enough
to hit the problem.
Author: Yura Sokolov
Reviewed-by: David Rowley, Ranier Vilela
Discussion: https://postgr.es/m/b1f7f32737c3438136f64b26f4852b96@postgrespro.ru
Backpatch-through: 10, where simplehash.h was added
Also "make reformat-dat-files".
The only change worthy of note is that pgindent messed up the formatting
of launcher.c's struct LogicalRepWorkerId, which led me to notice that
that struct wasn't used at all anymore, so I just took it out.
Comment fixes are applied on HEAD, and documentation improvements are
applied on back-branches where needed.
Author: Justin Pryzby
Discussion: https://postgr.es/m/20210408164008.GJ6592@telsasoft.com
Backpatch-through: 9.6
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu, Hou Zhijie
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
This removes "Add Result Cache executor node". It seems that something
weird is going on with the tracking of cache hits and misses as
highlighted by many buildfarm animals. It's not yet clear what the
problem is as other parts of the plan indicate that the cache did work
correctly, it's just the hits and misses that were being reported as 0.
This is especially a bad time to have the buildfarm so broken, so
reverting before too many more animals go red.
Discussion: https://postgr.es/m/CAApHDvq_hydhfovm4=izgWs+C5HqEeRScjMbOgbpC-jRAeK3Yw@mail.gmail.com
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
Previously simplehash.h only exposed a method to perform a hash table
delete using the hash table key. This meant that the delete function had
to perform a hash lookup in order to find the entry to delete. Here we
add a new function so that users of simplehash.h can perform a hash delete
directly using the entry pointer, thus saving the hash lookup.
An upcoming patch that uses simplehash.h already has performed the hash
lookup so already has the entry pointer. This change will allow the
code in that patch to perform the hash delete without the code in
simplehash.h having to perform an additional hash lookup.
Author: David Rowley
Reviewed-by: Andres Freund
Discussion: https://postgr.es/m/CAApHDvqFLXXge153WmPsjke5VGOSt7Ez0yD0c7eBXLfmWxs3Kw@mail.gmail.com
Masahiko Sawada
John Naylor
Andres Freund
Alexander Korotkov
Daniel Gustafsson
Nathan Bossart
Bruce Momjian
Peter Eisentraut
Jeff Davis
David Rowley
Michael Paquier
Tom Lane