Function rewrite_insert_assignments 

pub(crate) fn rewrite_insert_assignments(
    plan: LogicalPlan,
    query_ctx: &QueryContextRef,
    config: &ConfigOptions,
) -> Result<LogicalPlan>

Interprets strings assigned to timestamp columns at an INSERT boundary using the session timezone. plan is the assignment projection under a WriteOp::Insert.

DataFusion plans INSERT as a projection casting each source column to its target column type, and that cast reads a naive string as UTC. Arrow does apply a timezone when the cast target carries one, so the assignment is routed through Timestamp(unit, Some(tz)) and back. Stripping the timezone afterwards is value-preserving — arrow only shifts values in the opposite direction.

The source query is left untouched. Reinterpreting a value where it is produced would change what the source query means: pushing the conversion below a UNION’s DISTINCT, for instance, moves the dedup key from the raw strings to parsed instants and silently drops rows.

Why TypeCoercion runs here

The rewrite reads source types, and those are only settled once a UNION’s branch types have been reconciled: before coercion a union carries its loose schema (the first branch’s types), so a mixed SELECT 'string' UNION ALL SELECT CAST(.. AS TIMESTAMP) still looks like a string. Retargeting that cast would leave Timestamp(None) -> Timestamp(Some(tz)) behind once coercion retypes the union — the one direction in which arrow shifts the value instead of relabelling it.

Coercing here rather than deferring to the analyzer is forced by where an INSERT is still identifiable: exec_dml_statement strips the Dml node and executes its input, so by the time the analyzer runs, an assignment projection is indistinguishable from any other projection.

Explicit casts stay out of this: the SQL layer turns a user’s CAST(x AS TIMESTAMP) into an arrow_cast call, which only becomes an Expr::Cast in the optimizer’s SimplifyExpressions. Assignment casts are therefore the only Expr::Cast reaching a timestamp column here.

Reach

The emitted cast only carries its timezone where the expression is evaluated on this node. Substrait drops the timezone name when a plan is pushed down — it encodes any zoned timestamp as PrecisionTimestampTz and decodes it back as UTC — so a source reading from a table falls back to UTC, the behaviour it had before this rule existed. Sources that never leave this node (literals, VALUES, and UNIONs of them) keep the session timezone, and those are what an INSERT’s timestamp assignment is in practice.