Commits on Apr 20, 2021

1. sql: support EXPLAIN ANALYZE EXECUTE …

   We currently cannot EXPLAIN an EXECUTE statement (the syntax doesn't
   even allow it). EXECUTE is handled with special code at the top-level,
   so it's tricky to make it work with EXPLAIN. However, EXPLAIN ANALYZE
   Is also handled at the top level, before the EXECUTE handling, so that
   should just work.
   
   This change allows the parsing of such statements so we can use
   EXPLAIN ANALYZE EXECUTE. If EXPLAIN EXECUTE is attempted, we get a
   "not supported" error (this is less cryptic than the parsing error
   anyway).
   
   Release note: None

   

   RaduBerinde committed Apr 20, 2021
   Configuration menu

   • View commit details
   • Copy full SHA for 22e4d64
   • Browse repository at this point

   Copy the full SHA

   22e4d64 View commit details

   Browse the repository at this point in the history

2. opt: normalize In with single-element list to Eq …

   This commit adds a rule that normalizes `a IN (b)` to `a = b` (and
   similarly `NOT IN` to `!=`).
   
   This query confirms the equivalency holds even when NULLs are
   involved:
   
   ```
   WITH
     vals (v) AS (VALUES (1), (2), (NULL))
   SELECT
     v1.v, v2.v, v1.v NOT IN (v2.v,) AS in, v1.v != v2.v AS eq
   FROM
     vals AS v1, vals AS v2
   
      v   |  v   |  in   |  eq
   -------+------+-------+--------
        1 |    1 | false | false
        1 |    2 | true  | true
        1 | NULL | NULL  | NULL
        2 |    1 | true  | true
        2 |    2 | false | false
        2 | NULL | NULL  | NULL
     NULL |    1 | NULL  | NULL
     NULL |    2 | NULL  | NULL
     NULL | NULL | NULL  | NULL
   ```
   
   Release note: None

   

   RaduBerinde committed Apr 20, 2021
   Configuration menu

   • View commit details
   • Copy full SHA for 4c6e54e
   • Browse repository at this point

   Copy the full SHA

   4c6e54e View commit details

   Browse the repository at this point in the history

3. opt: introduce placeholder fast path …

   Currently running a query with placeholders happens as follows:
    - at preparation time, we build a memo with the normalized expression
      *with* placeholders.
    - at execution time, we copy the expression into a new memo,
      replacing placeholders with their value (AssignPlaceholders).
    - then we run exploration which performs the actual optimization.
   
   For trivial queries like KV reads, we do too much work during
   execution (profiles show it is 10-20% of the runtime for KV workloads
   with high read rates).
   
   This commit introduces the concept of a "placeholder fast path". The
   idea is that we can make specific checks for simple expressions and
   produce (at preparation time) a fully optimized expression (which
   still depends on placeholders). For this, we introduce a new operator,
   PlaceholderScan which is similar to a Scan except that it always scans
   one span with the same start and end key, and the key values are child
   scalar expressions (either constants or placeholders).
   
   We use this new operator only for simple SELECTs where the filters
   constrain a prefix of an index to constant values; in addition, the
   index must be covering and there must be only one such index. With
   these conditions, we know the optimal plan upfront.
   
   For now, the placeholder fast path transformation is Go code; if it
   gets more complicated, it should be switched to use optgen rules.
   
   A benchmark on my machine shows a kv95 workload going from 34kiops to
   38kiops with this change.
   
   Release note: None

   

   RaduBerinde committed Apr 20, 2021
   Configuration menu

   • View commit details
   • Copy full SHA for df67aa9
   • Browse repository at this point

   Copy the full SHA

   df67aa9 View commit details

   Browse the repository at this point in the history