JSON

Status: Native for both decode and encode, parity-pinned against Flink's own converters (JsonDecodeParityTest for decode), with the gaps enumerated below. Covers plain json and, on decode, the four CDC envelope dialects (debezium-json, canal-json, maxwell-json, ogg-json), which decode through the same column appenders as plain JSON. See Kafka for the connector-level table shapes (kafka, upsert-kafka) each format can be used in, and Configuration for the global native switches.

Decode (Kafka source)

Both timestamp-format.standard modes (SQL and ISO-8601) are native on every JSON-decoded path. Scalar coercion follows Flink's converters exactly: string-encoded numbers with trimming, Infinity/NaN/suffix floats, never-failing booleans, strict ISO_LOCAL_DATE dates, and integer/boolean/container echo under STRING. DECIMAL parses the exact raw literal with BigDecimal's HALF_UP-or-NULL rounding. ignore-parse-errors skips at Flink's own per-field granularity. A top-level JSON array fans out into one row per element, matching Flink's processArray (an empty array yields zero rows; a bad element fails the message in strict mode and drops alone under ignore-parse-errors). TIME columns parse SQL_TIME_FORMAT and reproduce Flink's silent sub-second discard (toSecondOfDay() * 1000, regardless of declared precision), including java.time's SMART hour-24-is-midnight resolution. VARBINARY reproduces Jackson's exact base64 read (whitespace between four-char groups, padding required, declared length not enforced), down to its corrupted-input drop granularity. The CDC dialects treat an array-rooted envelope as a corrupt message, as Flink does.

Supported column types (recursively over ROW/ARRAY/MAP/MULTISET): BOOLEAN, TINYINT, SMALLINT, INT, BIGINT, FLOAT, DOUBLE, CHAR/VARCHAR, DATE, TIME, TIMESTAMP, TIMESTAMP_LTZ, DECIMAL, VARBINARY.

Fallback condition	Why
fail-on-missing-field = true	Not modeled — a missing field decodes as null natively (Flink's default mode).
decode.json-parser.enabled = false	Switches Flink to its tree deserializer, whose coercion envelope differs from the parser path the native decode mirrors.
A column (or nested leaf) of type BINARY	Its fixed-size Arrow carriage can't hold arbitrary-length base64 without the length enforcement Flink's decode doesn't apply.
A column (or nested leaf) of an INTERVAL type	Outside the natively-converted type set.
A MAP/MULTISET key type outside CHAR/VARCHAR	Defensive only — Flink's own JSON format rejects a non-string map key at planning, so this can't reach substitution.

All Kafka startup modes (earliest/latest/group-offsets/timestamp/specific-offsets), topic lists, and topic-pattern are supported regardless of format, since discovery and offset resolution run in Flink's own reused enumerator.

Encode (Kafka sink)

Supported column types, recursively over ROW/ARRAY/MAP/MULTISET: BOOLEAN, TINYINT/SMALLINT/ INT/BIGINT, FLOAT/DOUBLE (see the spelling note below), CHAR/VARCHAR, BINARY/VARBINARY, DECIMAL, DATE (ISO_LOCAL_DATE, with +/- EXCEEDS_PAD years past 9999/below 0), TIME, TIMESTAMP, and TIMESTAMP_LTZ (SQL or ISO-8601). A null field inside a nested ROW follows encode.ignore-null-fields exactly as Flink's recursive converter does; array elements and map values keep explicit nulls regardless.

Map keys must be in the CHARACTER_STRING family (a MULTISET's element is its key) — Flink's own converter throws for anything else, so a non-string-keyed column declines and Flink raises its own error. Null map keys follow json.map-null-key.mode: DROP and LITERAL (with json.map-null-key.literal) reproduce Flink's bytes exactly; the default FAIL mode fails the record at runtime pointing at the option, as Flink's does, since a data-dependent failure can't gate at plan time. Duplicate map keys collapse the way Jackson's ObjectNode does — first position, last value — with one documented corner: a duplicate key whose value is a nested ROW under encode.ignore-null-fields merges field-by-field in Flink (an ObjectNode-reuse artifact) but takes the whole last value natively.

Option	Effect
encode.ignore-null-fields	Drops null fields (including a CDC envelope's null before/after key).
encode.decimal-as-plain-number	Keeps the column's declared scale; the default reproduces Jackson's stripTrailingZeros().toString(), scientific notation included.
json.map-null-key.mode	DROP / LITERAL / FAIL (default) — see above.
json.map-null-key.literal	The literal string substituted under LITERAL mode.

Each option set configures the format instance it belongs to, as in Flink: value options come from json.*/value.json.*, and upsert key options only from key.json.* — never inherited from the value format's settings.

Sink fallbacks specific to JSON

Fallback condition	Why
An out-of-range json.* option value	Flink's format factory raises its own validation error.
A field name that would need a JSON control-character escape	arrow-json spells field-name escapes lowercase where Jackson's are uppercase; value and map-key strings already escape natively in Jackson's exact form.
json.map-null-key.literal containing a line break	Not representable in the native encoder.
debezium-json.schema-include = true	Rejected by Flink's own sink factory.
A FLOAT/DOUBLE column when the runtime JDK float-spelling probe fails	See below.

The FLOAT/DOUBLE spelling probe. The native library ports the legacy (JDK ≤ 18) Double.toString/Float.toString algorithm — the parity target is JDK 17's spelling — but JDK 19 changed Double.toString to shortest-representation digits, which differ on roughly 0.3% of random doubles and 11% of random floats. At plan time the JVM spells a fixed corpus (deliberately containing values where the two algorithms disagree) and compares it against the native spelling; a mismatch keeps the column on the host rather than silently diverging. Decode is unaffected (parsing a JSON number is exact), and the reported reason is jdk float spelling mismatch (JDK 19+).

General sink-shape fallbacks that apply to every value format (an upsert-materialized sink, a keyed table, sink.parallelism on a changelog input, and so on) are covered on the Kafka page, not repeated here.