v20.2 vectorized updates #8009
v20.2 vectorized updates #8009
Conversation
|
This change is |
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/951f66c71342bdb61c6b1ca9961fa6a9289cbe80/ Edited pages: |
ericharmeling
force-pushed
the
20.2-vectorized
branch
from
951f66c to
67ef422
Compare
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/67ef42292441e2d3ff9dbcea60a035371d19bd74/ Edited pages: |
ericharmeling
force-pushed
the
20.2-vectorized
branch
from
67ef422 to
93a8df7
Compare
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/93a8df71fe7b83388d15d5778adb5c37dddcb28a/ Edited pages: |
ericharmeling
force-pushed
the
20.2-vectorized
branch
from
93a8df7 to
aa42a11
Compare
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/aa42a1138ecddd43b39386583fda0b3ce4c59f2a/ Edited pages: |
There was a problem hiding this comment.
Reviewed 11 of 11 files at r1.
Reviewable status:complete! 0 of 0 LGTMs obtained (waiting on @ericharmeling)
v20.2/vectorized-execution.md, line 19 at r1 (raw file):
Option | Description ----------|------------ `on` | Turns on vectorized execution for all queries on rows under the [`vectorize_row_count_threshold`](#setting-the-row-threshold-for-vectorized-execution) (1000 rows, by default).<br><br>**Default:** `vectorize=on`
s/rows under/rows over/.
v20.2/vectorized-execution.md, line 62 at r1 (raw file):
- Global [sorts](query-order.html) - [Window functions](window-functions.html)
I'd keep "window functions" here because they do require memory buffering (although we implement only several window functions in the vectorized engine, for most cases we will use old row-by-row processor that is embedded into the vectorized plan, but it'll still require buffering in memory). Also, I'd probably make the mention of "window functions" as the very last (I think it has lower priority than aggregations and joins).
v20.2/vectorized-execution.md, line 68 at r1 (raw file):
The vectorized engine does not support queries containing: - [Window functions](window-functions.html). See [tracking issue](https://github.com/cockroachdb/cockroach/issues/37040).
I'd call the support of window functions "limited" or "unoptimized" because if a query has a window function that we don't support but also other operations that we do support in the vectorized engine, we will still plan a vectorized flow with the caveat that the window function will be handled in the unoptimized row-by-row fashion, but other operations will be handled efficiently.
ericharmeling
changed the title
ericharmeling
force-pushed
the
20.2-vectorized
branch
from
aa42a11 to
a7713fa
Compare
There was a problem hiding this comment.
TFTR @yuzefovich !
Adding @asubiotto for another pass.
Reviewable status:
v20.2/vectorized-execution.md, line 19 at r1 (raw file):
Previously, yuzefovich wrote…
s/rows under/rows over/.
Done.
v20.2/vectorized-execution.md, line 62 at r1 (raw file):
Previously, yuzefovich wrote…
I'd keep "window functions" here because they do require memory buffering (although we implement only several window functions in the vectorized engine, for most cases we will use old row-by-row processor that is embedded into the vectorized plan, but it'll still require buffering in memory). Also, I'd probably make the mention of "window functions" as the very last (I think it has lower priority than aggregations and joins).
Done.
v20.2/vectorized-execution.md, line 68 at r1 (raw file):
Previously, yuzefovich wrote…
I'd call the support of window functions "limited" or "unoptimized" because if a query has a window function that we don't support but also other operations that we do support in the vectorized engine, we will still plan a vectorized flow with the caveat that the window function will be handled in the unoptimized row-by-row fashion, but other operations will be handled efficiently.
Done.
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/a7713face5749b11fda19f89e2024966d57ef43e/ Edited pages: |
There was a problem hiding this comment.
Reviewed 1 of 1 files at r2.
Reviewable status:
There was a problem hiding this comment.
Reviewable status:
v20.2/vectorized-execution.md, line 20 at r2 (raw file):
----------|------------ `on` | Turns on vectorized execution for all queries on rows over the [`vectorize_row_count_threshold`](#setting-the-row-threshold-for-vectorized-execution) (1000 rows, by default).<br><br>**Default:** `vectorize=on` `201auto` | Follows the [vectorized execution behavior of CockroachDB v20.1](../v20.1/vectorized-execution.html), instructing CockroachDB to use the vectorized execution engine on most queries that execute in memory, on [data types supported by the vectorized engine in CockroachDB v20.1](../v20.1/data-types.html), without the need to [spill intermediate results to disk](../v20.1/vectorized-execution.html#disk-spilling-operations).
This description is a bit unclear to me. I would change it to say something along the lines of "queries that use a constant amount of memory" and maybe remove the disk spilling mention
v20.2/vectorized-execution.md, line 70 at r2 (raw file):
- A join filtered with an [`ON` expression](joins.html#supported-join-conditions). See [tracking issue](https://github.com/cockroachdb/cockroach/issues/38018). - Any query containing constant `NULL` arguments, with the exception of the `IS` projection operators `IS NULL` and `IS NOT NULL`. For example, `SELECT x IS NOT NULL FROM t` is supported, but `SELECT x + NULL FROM t` returns an `unable to vectorize execution plan` error. See [tracking issue](https://github.com/cockroachdb/cockroach/issues/41001).
I think this NULL business is not true anymore because we now wrap the post-processing stages (also tried it out to confirm). cc @yuzefovich
ericharmeling
force-pushed
the
20.2-vectorized
branch
from
a7713fa to
0815d73
Compare
There was a problem hiding this comment.
TFTR @asubiotto.
I'm going to pass off to writer review now and then merge. We can always update docs ahead of/after the release if anything needs changing.
Reviewable status:
v20.2/vectorized-execution.md, line 20 at r2 (raw file):
Previously, asubiotto (Alfonso Subiotto Marqués) wrote…
This description is a bit unclear to me. I would change it to say something along the lines of "queries that use a constant amount of memory" and maybe remove the disk spilling mention
Done.
v20.2/vectorized-execution.md, line 70 at r2 (raw file):
Previously, asubiotto (Alfonso Subiotto Marqués) wrote…
I think this
NULLbusiness is not true anymore because we now wrap the post-processing stages (also tried it out to confirm). cc @yuzefovich
Removed this bullet.
|
Online preview: http://cockroach-docs-review.s3-website-us-east-1.amazonaws.com/0815d73a85144845b20bdb875b1f8dabc34427c0/ Edited pages: |
There was a problem hiding this comment.
Reviewable status:
There was a problem hiding this comment.
Reviewable status:
v20.2/vectorized-execution.md, line 70 at r2 (raw file):
Previously, ericharmeling (Eric Harmeling) wrote…
Removed this bullet.
The boundary what is "unsupported" is blurred because of wrapping - yes, queries with constant NULL arguments will have a wrapped row-by-row processor into the vectorized flow, so I agree with Alfonso, it's a good idea to remove this bullet.
Fixes #7845.
Fixes #7843.
Fixes #7844.
Fixes #7629.
Fixes #7618.
Fixes #7635.
Fixes #7842.
Fixes #7743.
Fixes #7628.
Fixes #8125.
Fixes #8141.
Fixes #8165.