Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
70722: sql: add assignment casts for INSERTs r=mgartner a=mgartner
#### sql: rename pkg/sql/sem/tree/casts.go to cast.go
Release note: None
#### tree: add OID->OID cast map
This commit adds a new map that describes valid casts from OID to OID.
It introduces three cast contexts: explicit casts, assignment casts, and
implicit casts. See the comments for CastContext and castMap for
details.
This map will enable us to properly follow Postgres's casting behavior.
Most immediately, it will allow us to support assignment casts.
Future work includes moving volatility in castMap. In the longer term,
cast functions can be moved into castMap as well.
Release note: None
#### sql: set "char" type width to 1
The `"char"` type is a special single-character type. This commit
adds a `types.QChar` with a width one. It removes the `types.MakeQChar`
function so that it is impossible to create `"char"` types with any
other width.
Release note: None
#### sql: add assignment casts for INSERTs
Casts in Postgres are performed in one of three contexts [1]:
1. An explicit context with `CAST(x AS T)` or `x::T`.
2. An assignment context performed implicitly during an INSERT,
UPSERT, or UPDATE.
3. An implicit context during the evaluation of an expression. For
example the DATE in `'2021-01-02'::DATE < now()` will be implicitly
cast to a TIMESTAMPTZ so that the values can be compared.
Not all casts can be performed in all contexts. Postgres's pg_cast table
lists valid casts and specifies the maximum cast context in which each
can be performed. A cast with a max context of explicit can only be
performed in an explicit context. A cast with an assignment max context
can be performed in an explicit context or an assignment context. A cast
with an implicit max context can be performed in all contexts.
Much to my personal disappointment and frustration, there are valid
casts that are not listed in Postgres's pg_cast table. These casts are
called "automatic I/O conversions" and they allow casting most types to
and from the string types: TEXT, VARCHAR, CHAR, NAME, and "char" [2].
We cannot determine these casts' maximum cast context from the pg_cast
table, so we rely on the documentation which states that conversions to
string types are assignment casts and conversions from string types are
explicit casts [3].
--
This commit implements assignment casts for INSERTs. Follow up work will
implement assignment casts for UPSERTs and UPDATEs.
A cast performed in an assignment context has slightly different
behavior than the same cast performed in an explicit context. In an
assignment context, the cast will error if the width of the value is too
large for the given type. In an explicit context, the value will be
truncated to match the width. The one exception is assignment casts to
the special "char" type which do truncate values.
To support different cast behaviors for different contexts, a new
built-in, `crdb_internal.assignment_cast` has been introduced. This
function takes two arguments: a value and a type. Because SQL
does not have first-class types, a type cannot be passed directly to the
built-in. Instead, a `NULL` cast to a type is used as a workaround,
similar to the `json_populate_record` built-in. For example, an integer
can be assignment-cast to a string with:
crdb_internal.assignment_cast(1::INT, NULL::STRING)
The optimizer is responsible for wrapping INSERT columns with the
assignment cast built-in function. If an insert column type `T1` is not
identical to the table's corresponding target column type `T2`, the
optimizer will check if there is a valid cast from `T1` to `T2` with a
maximum context that allows an assignment cast. If there is a such a
cast, a projection will wrap the column in the assignment cast built-in
function. If there is no such cast, a user error will be produced.
Some changes to prepared statement placeholder type inference were
required in order to better match Postgres's behavior (this is a
best-effort match thus far and there are still inconsistencies). Most
notably, widths and precision are no longer inferred for the types of
placeholders. The effect of this is that assignment casts will be
correctly added by the optimizer in order to make sure that values for
placeholders are correctly coerced to the target column type during
execution of a prepared insert.
The introduction of assignment casts fixes minor bugs and addresses some
inconsistencies with Postgres's behavior. In general, INSERTS now
successfully cast values to target table column types in more cases. As
one example, inserting a string into an integer column now succeeds:
CREATE TABLE t (i INT)
INSERT INTO t VALUES ('1'::STRING)
Prior to this commit there was logic that mimicked assignment casts, but
it was not correct. Bugs in the implementation caused incorrect behavior
when inserting into tables with computed columns. Most notably, a
computed column expression that referenced another column `c` was
evaluated with the value of `c` before the assignment cast was
performed. This resulted in incorrect values for computed columns in
some cases.
In addition, assignment casts make the special logic for rounding
decimal values in optbuilder obsolete. The builtin function
`crdb_internal.round_decimal_values` and related logic in optbuilder
will be removed once assignment casts are implemented for UPSERTs and
UPDATEs.
Fixes #69327
Fixes #69665
[1] https://www.postgresql.org/docs/current/typeconv.html
[2] https://www.postgresql.org/docs/13/catalog-pg-cast.html#CATALOG-PG-CAST
[3] https://www.postgresql.org/docs/13/sql-createcast.html#SQL-CREATECAST-NOTES
Release note (sql change): Implicit casts performed during INSERT
statements now more closely follow Postgres's behavior. Several minor
bugs related to these types of casts have been fixed.
70950: instancestorage: Add SQL test for sql_instances. r=knz a=rimadeodhar
This PR adds a unit test to verify that the sql_instances
table can be accessed through SQL API.
Release note: None
71412: sqlproxyccl: rework sqlproxy connection throttler r=JeffSwenson a=JeffSwenson
This change switches the sqlproxy connection throttling logic back to
exponential backoff. The tokenbucket approach was introduced by
PR #69041. There are a few behavior differences between this and the
original exponential backoff implementation.
1. The throttling logic is maintained per (ip, tenant) instead of per
(ip). Some platform as a service provides share a single outbound ip
address between multiple clients. These users would occasionaly see
throttling caused by a second user sharing their IP.
2. The throttling logic was triggered before there was an authentication
failure. It takes ~100ms-1000ms to authenticate with the tenant
process. Any requests that arrived after the first request, but
before it was processed, would trigger the throttle. Now, we only
trigger the throttle in response to an explict authorization error.
Release note: None
Co-authored-by: Marcus Gartner <[email protected]>
Co-authored-by: rimadeodhar <[email protected]>
Co-authored-by: Jeff <[email protected]>- Loading branch information
