sql/stats: add cluster setting to disable generation of stats forecasts

[michae2]

Add a new cluster setting sql.stats.forecasts.enabled that can be
used to disable generation of statistics forecasts in the stats cache.
Changing this setting will cause the stats cache to gradually evict and
reload all cached statistics as they are requested.

Fixes: #86350

Release justification: Low-risk update to new functionality.

Release note (sql change): Add a new cluster setting
sql.stats.forecasts.enabled which controls whether statistics
forecasts are generated by default for all tables. (Note that this is
different from the session setting optimizer_use_forecasts which
controls whether statistics forecasts are used when optimizing the
current query. If forecasting is disabled, then even if
optimizer_use_forecasts is true for a given query it won't have any
forecasts to use.)

[cockroach-teamcity]

This change is 

[michae2]

Note that the first commit is from #86834 and can be ignored.

[rytaft]

Reviewed 15 of 15 files at r1, 6 of 6 files at r2, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @msirek)

[michae2]

TFTR!

bors r=rytaft

[msirek]

Nice change. I have some comments. You may wish to hold off on the merge.

Reviewable status: complete! 0 of 0 LGTMs obtained (and 1 stale) (waiting on @michae2 and @rytaft)

---

-- commits line 15 at r1:
nit: maybe mention that both the session setting and cluster setting must be true for stats forecasts to be used, so it's clear in docs.

---

pkg/sql/opt/memo/memo.go line 192 at r1 (raw file):

		reorderJoinsLimit:                      int(evalCtx.SessionData().ReorderJoinsLimit),
		zigzagJoinEnabled:                      evalCtx.SessionData().ZigzagJoinEnabled,
		useForecasts:                           evalCtx.SessionData().OptimizerUseForecasts,

Since the stats in the stats cache now depends on both the setting of evalCtx.SessionData().OptimizerUseForecasts and cluster settingUseStatisticsForecasts, it looks like this line should be:

useForecasts:                           evalCtx.SessionData().OptimizerUseForecasts && forecastAllowed(clusterSettings)

---

pkg/sql/opt/memo/memo.go line 328 at r1 (raw file):

	if m.reorderJoinsLimit != int(evalCtx.SessionData().ReorderJoinsLimit) ||
		m.zigzagJoinEnabled != evalCtx.SessionData().ZigzagJoinEnabled ||
		m.useForecasts != evalCtx.SessionData().OptimizerUseForecasts ||

Similar to the above comment, if the cluster setting changes, we want the query plan to be regenerated instead of using the cached plan, so it should be something like:

m.useForecasts != (evalCtx.SessionData().OptimizerUseForecasts && forecastAllowed(clusterSettings) ||

---

pkg/sql/stats/stats_cache.go line 294 at r2 (raw file):

// forecastAllowed returns true if statistics forecasting is allowed for the
// given table.
func forecastAllowed(table catalog.TableDescriptor, clusterSettings *cluster.Settings) bool {

Remove parameter table. It is unused.

[craig]

Build failed (retrying...):

• Bazel Essential CI (Cockroach)

[michae2]

Oh, thank you for taking a look! I'll hold off while we review.

bors r-

I think some of the things you point out will make more sense when the next PR, adding a per-table setting to enable or disable forecasts, is opened.

Reviewable status: complete! 0 of 0 LGTMs obtained (and 1 stale) (waiting on @msirek and @rytaft)

---

-- commits line 15 at r1:

Previously, msirek (Mark Sirek) wrote…

nit: maybe mention that both the session setting and cluster setting must be true for stats forecasts to be used, so it's clear in docs.

Good point!

---

pkg/sql/opt/memo/memo.go line 192 at r1 (raw file):

Previously, msirek (Mark Sirek) wrote…

Since the stats in the stats cache now depends on both the setting of evalCtx.SessionData().OptimizerUseForecasts and cluster settingUseStatisticsForecasts, it looks like this line should be:

useForecasts:                           evalCtx.SessionData().OptimizerUseForecasts && forecastAllowed(clusterSettings)

This useForecasts is similar to useHistograms and useMultiColStats: it represents whether we want this particular query to use forecasts, regardless of whether they are available or not. The same is true of useHistograms and useMultiColStats.

In the next PR, forecastAllowed will be a per-table function, so it won't make sense to call here (in a per-query setting).

---

pkg/sql/opt/memo/memo.go line 328 at r1 (raw file):

Previously, msirek (Mark Sirek) wrote…

Similar to the above comment, if the cluster setting changes, we want the query plan to be regenerated instead of using the cached plan, so it should be something like:

m.useForecasts != (evalCtx.SessionData().OptimizerUseForecasts && forecastAllowed(clusterSettings) ||

Invalidation when table statistics change is already handled by the call to m.Metadata().CheckDependencies below. We only need to invalidate the query plan when statistics actually change (which might not necessarily happen when enabling or disabling forecasts) and this is handled by the call to optTable.Equals inside CheckDependencies.

---

pkg/sql/stats/stats_cache.go line 294 at r2 (raw file):

Previously, msirek (Mark Sirek) wrote…

Remove parameter table. It is unused.

table will be used by the next PR, adding a per-table setting to fix #86353 (which will be opened shortly). I included it here to make sure I was setting everything up to also work with a per-table setting.

[craig]

Canceled.

[msirek]

Reviewed 7 of 7 files at r3.
Reviewable status: complete! 1 of 0 LGTMs obtained (and 1 stale) (waiting on @michae2 and @msirek)

---

pkg/sql/opt/memo/memo.go line 192 at r1 (raw file):

Previously, michae2 (Michael Erickson) wrote…

This useForecasts is similar to useHistograms and useMultiColStats: it represents whether we want this particular query to use forecasts, regardless of whether they are available or not. The same is true of useHistograms and useMultiColStats.

In the next PR, forecastAllowed will be a per-table function, so it won't make sense to call here (in a per-query setting).

I see. Thanks for the explanation.

---

pkg/sql/opt/memo/memo.go line 328 at r1 (raw file):

Previously, michae2 (Michael Erickson) wrote…

Invalidation when table statistics change is already handled by the call to m.Metadata().CheckDependencies below. We only need to invalidate the query plan when statistics actually change (which might not necessarily happen when enabling or disabling forecasts) and this is handled by the call to optTable.Equals inside CheckDependencies.

OK, I see. It looks like this following code would handle it:

cockroach/pkg/sql/opt_catalog.go

Lines 1084 to 1092 in 1540ea8

	// Verify the stats are identical.
	if len(ot.stats) != len(otherTable.stats) {
	return false
	}
	for i := range ot.stats {
	if !ot.stats[i].equals(&otherTable.stats[i]) {
	return false
	}
	}

When the cluster setting is changed, dataSourceForTable would get a different length list of table stats.

---

pkg/sql/stats/stats_cache.go line 294 at r2 (raw file):

Previously, michae2 (Michael Erickson) wrote…

table will be used by the next PR, adding a per-table setting to fix #86353 (which will be opened shortly). I included it here to make sure I was setting everything up to also work with a per-table setting.

OK

[michae2]

TFTRs!

bors r=rytaft,msirek

[craig]

Build failed:

• Bazel Essential CI (Cockroach)

[michae2]

Trying again.

bors r=rytaft,msirek

[craig]

Build succeeded:

• Bazel Essential CI (Cockroach)