Add EdgeDB query performance statistics

[fantix]

Refs #7725, this is a quick'n'dirty first attempt the first implementation with a standalone extension that manipulates the PG query source text (and optionally the queryID to map to the cache key). The upper side is that the extension is kinda clean without patching PG codebase (easier to port), the down side is that we lose the raw SQL stats data for all.

• Fix tests
• Deal with SQL user permission issue
• Fix missing hooks
• Add EdgeQL view
• Add backend capabilities
• Handle multitenancy correctly
• Play nice with persistent cache
• Complete the reset function parameters
• Count SQL queries in
• Add columns that reflect compilation settings
• Track EdgeQL scripts correctly
• Fix CI builds
• Add test

Sample results:

_localdev:main> select sys::QueryStats { * } filter .query not like '%QueryStats%' and .query not like '%sys::%' limit 2;
{
  sys::QueryStats {
    computed_fields: [],
    builtin: false,
    internal: false,
    name: '4ed18524-e5ee-6288-a0b0-9a171bcedd98',
    id: 4ed18524-e5ee-6288-a0b0-9a171bcedd98,
    query: 'SELECT 89',
    query_id: 5679466998613828232,
    query_type: SQL,
    compilation_config: Json("{\"allow_user_specified_id\": [\"false\"]}"),
    protocol_version: (major := 3, minor := 0),
    default_namespace: 'public',
    namespace_aliases: {},
    migration_name: 'initial',
    output_format: {},
    expect_one: {},
    implicit_limit: {},
    inline_typeids: {},
    inline_typenames: {},
    inline_objectids: {},
    plans: 0,
    total_plan_time: <duration>'0:00:00',
    min_plan_time: <duration>'0:00:00',
    max_plan_time: <duration>'0:00:00',
    mean_plan_time: <duration>'0:00:00',
    stddev_plan_time: <duration>'0:00:00',
    calls: 1,
    total_exec_time: <duration>'0:00:00.000001',
    min_exec_time: <duration>'0:00:00.000001',
    max_exec_time: <duration>'0:00:00.000001',
    mean_exec_time: <duration>'0:00:00.000001',
    stddev_exec_time: <duration>'0:00:00',
    rows: 1,
  },
  sys::QueryStats {
    computed_fields: [],
    builtin: false,
    internal: false,
    name: 'bcecb044-dd0b-af72-d014-345032436dfc',
    id: bcecb044-dd0b-af72-d014-345032436dfc,
    query: 'select
    (BB, (<__std__::int64>$0 + <__std__::int64>$1))',
    query_id: -4833294490262589582,
    query_type: EdgeQL,
    compilation_config: Json("{\"allow_bare_ddl\": \"AlwaysAllow\", \"query_cache_mode\": \"Default\", \"__internal_testmode\": false, \"store_migration_sdl\": \"NeverStore\", \"force_database_error\": \"false\", \"apply_access_policies\": true, \"allow_dml_in_functions\": false, \"allow_user_specified_id\": false, \"__internal_query_reflschema\": false, \"__internal_no_apply_query_rewrites\": false}"),
    protocol_version: (major := 2, minor := 0),
    default_namespace: 'bb',
    namespace_aliases: {},
    migration_name: 'm1ndk7eqxayvb5wdfkuv7kslyx2wnhr3e3afcmakqtgo4hjcxwspka',
    output_format: BINARY,
    expect_one: false,
    implicit_limit: 0,
    inline_typeids: false,
    inline_typenames: false,
    inline_objectids: true,
    plans: 6,
    total_plan_time: <duration>'0:00:00.001099',
    min_plan_time: <duration>'0:00:00.000086',
    max_plan_time: <duration>'0:00:00.000227',
    mean_plan_time: <duration>'0:00:00.000183',
    stddev_plan_time: <duration>'0:00:00.000048',
    calls: 10,
    total_exec_time: <duration>'0:00:00.000118',
    min_exec_time: <duration>'0:00:00.000007',
    max_exec_time: <duration>'0:00:00.000019',
    mean_exec_time: <duration>'0:00:00.000012',
    stddev_exec_time: <duration>'0:00:00.000004',
    rows: 0,
  },
}

[elprans]

the down side is that we lose the raw SQL stats data for all.

But we don't have to, do we? If we don't recognize a query as being a compiler product we can avoid modifying its state.

[1st1]

What version of C will this be compiled with? Because you could just do this in C99:

Suggested change

	EdbStmtInfoSemState state;
	EdbStmtInfoSemState state = {0};

[fantix]

Yeah, mostly assuming C99 in #7869 now

[msullivan]

Can we make sure nothing secret goes in also?

[fantix]

I think the only compilation setting that may include free text is force_database_error, which shouldn't have secrets. I can also add a double-safe to filter out extension configs in #7869.

[msullivan]

Do we want whether we collect this to also depend on a config flag? This seems like a lot of extra stuff to store unconditionally.
(Curious about @elprans's opinion on this)

[fantix]

Answered in the meeting: we do need a config to turn the tracking on/off.

Follow-up question: do we want to make this config affects_compilation (thus controlling the compiled SQL with or without the magic JSON-comment prefix, but impacting query cache), or a runtime session config that the extension can read and behave (but we end up with prefixed SQLs all the time)?