ChDriver.
Copy Markdown
Decoders for ClickHouse's wrapper and compound column types: Nullable(T),
Array(T), Map(K, V), LowCardinality(T), Variant(T1, ..., Tn), and
Decimal(P, S).
Dispatched from ChDriver.Protocol.NativeBlock's decode_column_data/3,
which these functions recurse back into for their inner type(s) — that's
what lets deeply nested types like Array(Nullable(String)) or
Map(String, Array(UInt32)) decode correctly without any special-casing.
Map(K, V) values decode to plain Elixir maps; Array(T) and
LowCardinality(T) decode to lists; Nullable(T) decodes to the inner
value or nil; Variant(...) decodes to whichever alternative each row
selected (or nil); Decimal(P, S) decodes to a Decimal.t(). See
ARCHITECTURE.md for the wire-level byte layouts.
Each decoder takes and returns the hoisted serialization-prefix list —
see ChDriver.Protocol.NativeBlock's moduledoc for why prefixes are read
up front rather than inline.
Summary
Functions
Decodes a Variant(T1, ..., Tn) column: one discriminator byte per row
(the 0-based index of the alternative that row holds, in the type name's
order, or 255 for no value), followed by one contiguous sub-column per
alternative, in alternative order, holding only the rows that selected
it.
Splits values (the flattened element array) back into per-row lists
using Array(T)'s cumulative offsets, e.g. values = [1, 2, 3, 4, 5]
and offsets = [2, 2, 5] (row 0 has 2 elements, row 1 has 0, row 2 has
Functions
Decodes a Variant(T1, ..., Tn) column: one discriminator byte per row
(the 0-based index of the alternative that row holds, in the type name's
order, or 255 for no value), followed by one contiguous sub-column per
alternative, in alternative order, holding only the rows that selected
it.
An alternative no row selected contributes zero bytes, so the discriminators have to be counted before any sub-column can be read — which is exactly why this can't be a streaming per-row decode.
Variant's own 8-byte discriminator-mode prefix was already consumed by
ChDriver.Protocol.NativeBlock.decode_prefixes/2, so it's popped off
prefixes here rather than read from binary.
Splits values (the flattened element array) back into per-row lists
using Array(T)'s cumulative offsets, e.g. values = [1, 2, 3, 4, 5]
and offsets = [2, 2, 5] (row 0 has 2 elements, row 1 has 0, row 2 has
- splits into
[[1, 2], [], [3, 4, 5]].