Next steps on pg_stat_statements normalisation

From: Peter Geoghegan <peter(at)2ndquadrant(dot)com>
To: PG Hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Next steps on pg_stat_statements normalisation
Date: 2012-01-22 05:30:02
Message-ID: CAEYLb_W4oUZ+qE6t8Bm_YpXVU=33i6MvxXt43sHpsZGVw5TD=A@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

So, having received feedback from Tom and others in relation to this
patch, I would like to state how I think I should go about addressing
various concerns to ensure that a revision of the patch gets into the
9.2 release. As I've said time and again, I think that it is very
important that we have this, sooner rather than later.

The main objection made in relation to my patch was to the overhead of
the admittedly invasive (though purely mechanical) changes to the
grammar to record the length of all tokens, so that Const tokens could
subsequently have their length known in a principled fashion. While I
believe that you'd be extremely hard pushed to demonstrate any
regression at all from these extra cycles, it's also true that given
the starting position of tokens (a value already exposed for the
purposes of producing error messages outside the parser that reference
particular tokens with a caret to their position in the query string,
corresponding to various nodes) there's no reason to think that it
should not be possible to calculate the total length on-the-fly from
within pg_stat_statements.

The syntax for constants is sufficiently simple that I think that a
good set of regression tests should make this entirely practicable,
covering all permutations of relevant factors affecting how the
implementation should act, including for example
standard_conforming_strings. There's no reason to think that the SQL
syntax rules for constants should need to change very frequently, or
even at all, so we should be fine with just knowing the starting
position. It's quicker and easier to do it this way than to argue the
case for my original implementation, so that's what I'll do. Whatever
overhead this independent, pg_stat_statements-internal const parsing
may impose, it will at least only be paid once per query, when we
first require a canonicalised representation of the query for the
pg_stat_statements view.

I have, in the past, spoken against the idea of parsing query strings
on the fly (an objection proven to be well-founded by various third
party tools), but this is something quite different - by starting from
a position that is known to be where a Const token starts, we need
only concern ourselves with constant syntax, a much more limited and
stable thing to have to pin down.

I imagined that by plugging into Planner_hook, and hashing the
post-rewrite query tree immediately before it was passed to the
planner, that there was a certain amount of merit in recognising as
equivalent queries that would definitely result in the same plan, all
other things, and selectivity estimates of constants, being equal.
However, since some of that logic, such as the logic through which
different though equivalent join syntaxes are normalised, actually
occurs in the planner, it was judged that there was an undesirable
inconsistency. I also felt that it was useful that the hashing occur
on the post-rewrite "query proper", but various examples illustrated
that that might be problematic, such as the inability for the
implementation to differentiate constants in the original query string
from, for example, constants appearing in view definitions.

Tom wanted to hash plans. I think that this view was partially
informed by concerns about co-ordination of hooks. While I had worked
around those problems, the solution that I'd proposed wasn't very
pretty, and it wasn't that hard to imagine someone finding a reason to
need to break it. While I think there might be a lot to be said for
hashing plans, that isn't a natural adjunct to pg_stat_statements -
the user expects to see exactly one entry for what they consider to be
the same query, and I think you'd have a hard time convincing many
people that it's okay to have a large number of entries for what they
thought was the same query, just reflecting the fact that a different
plan was used for each of several entries with the same query string,
based on factors they didn't consider essential to the query that may
be quite transitory. Some more marginal plans for the same query could
be missing entirely from the fixed-size hash table. It wouldn't have
been a very useful way of normalising queries, even if it did more
accurately reflect where execution costs came from, so I am dead set
against doing this for normalisation. However, it might be that we can
add even more value in the 9.3 release by hashing plans as well, to do
entirely more interesting things, but we need to have a canonicalised,
normalised representation of queries first - that's obviously how
people expect a tool like this to work, and it's difficult to imagine
a practical work-flow for isolating poorly performing queries/plans
that doesn't stem from that.

Balancing all of these concerns, here's a basic outline of what I
think that the way forward is:

* Add a hook to core that allows modules to hook into the call to
parse_analyze() and parse_analyze_varparams(). There'd probably be two
single-line plugin functions for each, so that they're both
implemented within a function that does all the real work.

* Add an int64 value, query_id, to both Query and PlannedStmt structs,
that are just blindly copied from one to the other by the core system.
No need to worry about about the query tree and plan not being
sufficiently coupled now, as was previously a concern with prepared
queries, as that value will persist.

* Within pg_stat_statements, hook into parse_analyze(). Hash the post
analysis Query returned by parse_analyze() (and not within a planner
hook, which this proposed new revision no longer has).

Now, I know that Tom felt that equivalent syntaxes shouldn't be
normalized, or at least that this was trivial, and this approach would
still do some measure of that. However, having a big case statement
like the one in transformStmt() to act on the pre-analysis parse tree
(which is always referenced through a Node pointer, as it is
polymorphic) seems like an awful lot more code for little benefit. I'm
not even sure right now if doing that would avoid this implementation
artefact. In any case, ISTM that the solution is to simply not
document the precise behaviour of normalisation, so the user has no
reasonable basis for expecting, for example, that a cross join should
be treated the same as when the comma syntax is used. Is it really
defensible to extrapolate that that should happen from the observation
that it works when noise words like "outer" in "left outer join" are
omitted? I think not.

* Copy the hash to the Query struct's query_id, so that the plan that
our executor hook will later see will bear that same query_id.

* Track the position of constants within the query string when walking
the post analysis query tree if we don't have a hash table entry for
it already - do update the constants if we have constants for the
query_id already, but no hashtable entry yet. Keep this bookkeeping
infromation in palloc()'d memory, using some kind of associative data
structure that associates the list of constant positions with their
query_id. All of this avoids using the wrong set of constants with an
equivalent query that has a different amount of whitespace or
something like that.

* Within the existing executor hook, look for the query in the hash
table based on the PlannedStmt's query_id, trusting that the core
system has copied that value from the original Query tree, when the
query was originally parsed, potentially quite a while ago for
prepared queries. If it's recognised, just update the stats. If it is
not, make a new entry. This entry will require a canonicalised query
string; generate one from the query string the executor plugin already
sees, plus a list of constant starting positions previously computed
for this query_id/query string, figuring out the length of the
constant by simply knowing the rules for constants and applying them,
using the principles I've already described. Free the memory in the
bookkeeping structure allocated for this entry, safe in the knowledge
that the hash table entry will prevent that from being updated/needed,
until it is evicted from the hash, at which point if it is seen again
constant positions needs to be figured out again (that new
canonicalised query string might have more whitespace than before,
obviously).

Thoughts?

--
Peter Geoghegan       http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Jeff Janes 2012-01-22 05:48:19 Re: Avoid FK validations for no-rewrite ALTER TABLE ALTER TYPE
Previous Message Noah Misch 2012-01-22 05:23:55 Re: Second thoughts on CheckIndexCompatible() vs. operator families