[RFC] Plan cache

[czpmango]

Summary

Cache historical execution plans to avoid replanning for the same/similar query.

Motivation

Avoid replanning to reduce the overhead of the planner and optimizer.

Usage explanation

Invisible to the user.

Design explanation

• two-level cache
  • string cache
    Hash the query statement.
  • AST cache (match (v{age:18}) <=> match (v) where v.age=18)
    Hash the normalized AST.
• configs
  • cache size
  • replan_interval : validity period of cached plan
  • statistics_divergence_threshold : tolerance threshold for statistical changes
• LFU
  Cache elimination strategy.

Note: schema changes(such as index/constraint creation/removal) will flush the entire query plan cache.

Rationale and alternatives

Drawbacks

This may be a trade-off in the CBO scenario, which mean It may cause performance degradation of the execution plan.

Prior art

Unresolved questions

• can't apply to CBO scenarios for nebula
• ...

Future possibilities

• supports CBO scenarios
• ...

[Shylock-Hg]

Do you consider the explicit usage like the prepare statement in sql, e.g. https://www.postgresql.org/docs/9.4/sql-prepare.html ?

[Shylock-Hg]

I think you could cache the generated plan before CBO stage, and do CBO again for plan which instantiated from cache.

[czpmango]

cache the generated plan before CBO stage, and

Setting statistics_divergence_threshold to 0 has the same effect.

[czpmango]

Do you consider the explicit usage like the prepare statement in sql, e.g. https://www.postgresql.org/docs/9.4/sql-prepare.html ?

Just the same thing. Implicit usage is more user-friendly, I think.

[Shylock-Hg]

Do you consider the explicit usage like the prepare statement in sql, e.g. https://www.postgresql.org/docs/9.4/sql-prepare.html ?

Just the same thing. Implicit usage is more user-friendly, I think.

Yes, I think could provide both. The implicit mode is enable by flag, the explicit is enable by query directly.

[czpmango]

Do you consider the explicit usage like the prepare statement in sql, e.g. https://www.postgresql.org/docs/9.4/sql-prepare.html ?

Just the same thing. Implicit usage is more user-friendly, I think.

Yes, I think could provide both. The implicit mode is enable by flag, the explicit is enable by query directly.

It can be considered, but will increase the complexity of the syntax.
This RFC mainly refers to the implementation of neo4j.

[czpmango]

Out of date.