Обсуждение: Re: caching query results
> Ed Loehr wrote: > > > What would persuasive numbers look like? > > > > As a novice, I'd guess key questions would seem to be... > > > > How often is a query run in which the results are identical to previous > > invocations of that query? > > > > Typically send/recv rates vs. typical query planning/exec time? > > So wouldn't you get most of what you want if you could store a query plan? This should wait until after the proposed querytree overhaul we have in mind. I already discussed it with Tom Lane. The idea goes like this: After the overhaul, the rewriter is a very simple and fast step. So we could hook into the rewriter, who builds for EVERY query kinda key based on the nodes, relations and functions that appear in the querytree. These keys could be managed in a shared LRU table, and if the same key appears a number of times (0-n), it'sentire querytree + plan (after planning) will be saved into the shared mem. At a subsequent occurence, thequerycache will look closer onto the two trees, if they are really identically WRT all the nodes. Ifonly constant values have changed, the already known plan could be reused. Postmaster startup options for tuning that come into mind then are querycache memsize, minimum # of appearencebefore caching, maximum lifetime or # usage of a plan and the like. So setting the memsize to zero willcompletely disable and fallback to current behavior. After that, the entire parsing is still done for every query (so application level controlled query cacheing isstill another thing to care for). We would only be able to skip the planner/optimizer step. The question thereforeis how much of the entire processing time for a query can be saved if replacing this step by some sharedmemory overhead. I'm not sure if this is worth the entire efford at all, and we can only judge after the querytreeoverhaul is done. Then again, improving the query optimizer directly, so he's able to make better join orderdecisions faster, might be the way to go. Jan -- #======================================================================# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #========================================= wieck@debis.com (Jan Wieck) #
On Mon, 3 Apr 2000, Jan Wieck wrote:
> This should wait until after the proposed querytree overhaul
> we have in mind. I already discussed it with Tom Lane. The
> idea goes like this:
This is very interesting discussion for me, I have prepared code for
7.1? with PREPARE/EXECUTE commands and SPI changes for query cache. This
features allow users define cache entry (via SPI or PREPARE). But it is
already discussed before now. A validity of plans is a problem.
> After the overhaul, the rewriter is a very simple and fast
> step. So we could hook into the rewriter, who builds for
> EVERY query kinda key based on the nodes, relations and
> functions that appear in the querytree.
It is good idea. What exactly is a key? If I good understand this key
is for query identification only. Right?
> These keys could be managed in a shared LRU table, and if the
My current code is based on HASH table with keys and query&plan is
saved in special for a plan created MemoryContext (it is good for
a example SPI_freeplan()).
> same key appears a number of times (0-n), it's entire
> querytree + plan (after planning) will be saved into the
> shared mem.
Here I not understend. Why is here any time checking?
> At a subsequent occurence, the querycache will
> look closer onto the two trees, if they are really
> identically WRT all the nodes. If only constant values have
> changed, the already known plan could be reused.
IMHO users can use PREPARE / EXECUTE for same query. Suggested idea is
really good if this query cache will in shared memory and more backends
can use it.
Good. It is solution for 'known-query' and allow it skip any steps in the
query path. But we still not have any idea for cached plans validity. What
if user changes oid for any operator, drop column (etc)?
Or I anything overlook?
> We would only be able to skip the
> planner/optimizer step.
Instead a PREPARE/EXECUTE which allow skip all in the query path (only
executor is called. But it works for user defined query not for every
query.
> The question therefore is how much of
> the entire processing time for a query can be saved if
> replacing this step by some shared memory overhead. I'm not
> sure if this is worth the entire efford at all, and we can
> only judge after the querytree overhaul is done. Then again,
> improving the query optimizer directly, so he's able to make
> better join order decisions faster, might be the way to go.
Is really sure that this will faster? (it must create key for nodes,
search same query in any table (cache), copy new query&plan to cache
..etc.)
Karel
Karel Zak wrote:
> On Mon, 3 Apr 2000, Jan Wieck wrote:
>
> It is good idea. What exactly is a key? If I good understand this key
> is for query identification only. Right?
Right. Imagine a querytree (after overhaul) that looks like this:
+------+ | SORT | +------+ ^
| +-----------------------------+ | JOIN | |
atts:rel1.att1, rel2.att2 | | qual: rel1.att2 = rel2.att1 | +-----------------------------+
^ ^ | | +------------------+ +------------------+ |
SCAN | | SCAN | | rel: rel1 | | rel: rel2 | | atts: att1, att2 | |
atts:att1, att2 | +------------------+ +------------------+
which is a node structure describing a query of:
SELECT rel1.att1, rel2.att2 FROM rel1, rel2 WHERE rel1.att2 = rel2.att1;
The "key" identifying this querytree now could look like
SORT(JOIN(1.1,2.2;SCAN(78991;1,2),SCAN(78995;1,2);))
78991 and 78995 are the OIDs of rel1 and rel2. So the key is a very simplified description of what the query
does,and maybe the qualification should be included too. But it's enough to find a few candidates to look at
closeron the node level out of hundreds of cached plans.
> > These keys could be managed in a shared LRU table, and if the
>
> My current code is based on HASH table with keys and query&plan is
> saved in special for a plan created MemoryContext (it is good for
> a example SPI_freeplan()).
IIRC our hash table code insists on using global, per backend memory. I thought about managing the entire
querycachewith a new type of memory context, using different routines for palloc()/pfree(), working in a
sharedmemory area only and eventually freeing longest unused plans until allocation fits. Let's see if
usinghash tables here would be easy or not.
> > same key appears a number of times (0-n), it's entire
> > querytree + plan (after planning) will be saved into the
> > shared mem.
>
> Here I not understend. Why is here any time checking?
There's not that big of a win if you do all the shared memory overhead for any query at it's first
occurance.Such a generic query cache only makes sense for queries that occur often. So at it's first to n-th
occurancewe only count by key and after we know that it's one of these again'n'again thingies, we pay the cache
overhead.
Also I think, keeping the number of exclusive cache locks (for writing) as small as possible would be a good
idea WRT concurrency.
> IMHO users can use PREPARE / EXECUTE for same query. Suggested idea is
> really good if this query cache will in shared memory and more backends
> can use it.
Exactly that's the idea. And since the postmaster will hold the shared memory as it does for the block and
syscache, it'll survive even times of no DB activity.
> Good. It is solution for 'known-query' and allow it skip any steps in the
> query path. But we still not have any idea for cached plans validity. What
> if user changes oid for any operator, drop column (etc)?
That's why the key is only good to find "candidates". The cacheing has to look very close to the nodes in the
tree and maybe compare down to pg_attribute oid's etc. to decide if it's really the same query or not.
> Is really sure that this will faster? (it must create key for nodes,
> search same query in any table (cache), copy new query&plan to cache
> ..etc.)
Only some timing code put into backends in various real world databases can tell how much of the entire
processingtime is spent in the optimizer.
And I'd not be surprised if most of the time is already spent during the parse step, which we cannot skip
by this technique.
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#========================================= wieck@debis.com (Jan Wieck) #
On Tue, 4 Apr 2000, Jan Wieck wrote:
> Right. Imagine a querytree (after overhaul) that looks like
> this:
>
> +------+
> | SORT |
> +------+
> ^
> |
> +-----------------------------+
> | JOIN |
> | atts: rel1.att1, rel2.att2 |
> | qual: rel1.att2 = rel2.att1 |
> +-----------------------------+
> ^ ^
> | |
> +------------------+ +------------------+
> | SCAN | | SCAN |
> | rel: rel1 | | rel: rel2 |
> | atts: att1, att2 | | atts: att1, att2 |
> +------------------+ +------------------+
>
> which is a node structure describing a query of:
>
> SELECT rel1.att1, rel2.att2 FROM rel1, rel2
> WHERE rel1.att2 = rel2.att1;
>
> The "key" identifying this querytree now could look like
>
> SORT(JOIN(1.1,2.2;SCAN(78991;1,2),SCAN(78995;1,2);))
The nice picture. Thanks, I undestend you now. A question is where create
this key - create a specific function that look at to querytree and return
key or calculate it during statement transformation (analyze.c ..etc.).
Or is any other idea?
> > > These keys could be managed in a shared LRU table, and if the
> >
> > My current code is based on HASH table with keys and query&plan is
> > saved in special for a plan created MemoryContext (it is good for
> > a example SPI_freeplan()).
I thought about it, and what for SPI and PREPARE/EXECUTE query cache
use shared memory too? I'm vote for one query cache in postgresql. IMHO
is not good create a specific cache for SPI_saveplan()+PREPARE and second
for your suggested query cache.
If plans saved via SPI (under defined key - 'by_key' interface) will shared
under all backends a lot of features will faster (FK, PLangs ..etc) and
shared plans cached via PREPARE will persistent across more connetions.
(Some web developers will happy :-)
But I not sure with this...
> IIRC our hash table code insists on using global, per backend
> memory. I thought about managing the entire querycache with
> a new type of memory context, using different routines for
> palloc()/pfree(), working in a shared memory area only and
> eventually freeing longest unused plans until allocation
> fits. Let's see if using hash tables here would be easy or
> not.
I look at the current shmem routines - create specific space and hash
table for a query cache is not a problem, hash routines are prepared
for usage under shmem. The current lock management code is very simular.
With hash is not a problem here.
A problem is how store (copy) query & plan tree to this (shared) memory.
The current copyObject() is based on palloc()/pfree() and as you said
we haven't memory management routines (like palloc()) that working in
shmem.
Would be nice have MemoryContext routines for shmem - example
CreateGlobalMemory_in_shmem() and palloc() that knows work with this
specific context. It is a dream?
A solution is convert query & plan tree to string (like pg_rewrite (views))
and save to cache this string, (and what a speed during (vice versa) parsing?).
IMHO for this solution we not need a hash table, we can use a standard system
table and a syscache.
But more nice is variant with non-string and full plan-tree-structs in a
shmem.
> > Good. It is solution for 'known-query' and allow it skip any steps in the
> > query path. But we still not have any idea for cached plans validity. What
> > if user changes oid for any operator, drop column (etc)?
>
> That's why the key is only good to find "candidates". The
> cacheing has to look very close to the nodes in the tree and
> maybe compare down to pg_attribute oid's etc. to decide if
> it's really the same query or not.
OK.
Karel