mjsax commented on code in PR #22165:
URL: https://github.com/apache/kafka/pull/22165#discussion_r3654754480
##########
streams/src/main/java/org/apache/kafka/streams/state/internals/InMemoryTimeOrderedKeyValueChangeBuffer.java:
##########
@@ -465,28 +512,136 @@ private byte[] internalPriorValueForBuffered(final Bytes
key) {
}
}
+ // The legacy V0/V1 restore paths feed the currently-buffered prior value
back in as plain value
+ // bytes, so unwrap it when the in-memory encoding is the headers-aware
format.
+ private byte[] plainPriorValueForBuffered(final Bytes key) {
+ final byte[] priorValue = internalPriorValueForBuffered(key);
+ return storeHeaders ? Utils.rawPlainValue(priorValue) : priorValue;
+ }
+
+ // Normalizes a restored buffer value so its value parts match the
encoding currently used
+ // in-memory (ValueTimestampHeaders blobs when header stores are enabled,
plain values otherwise),
+ // regardless of the changelog version it was read from.
+ private BufferValue normalizeEncoding(final BufferValue value, final
boolean restoredWithHeaders) {
+ if (storeHeaders == restoredWithHeaders) {
+ return value;
+ }
+ if (storeHeaders) {
+ // Plain -> ValueTimestampHeaders. The original headers/timestamp
of legacy records are
+ // unknown, so we use empty headers and the record-context
timestamp.
+ final long timestamp = value.context().timestamp();
+ return new BufferValue(
+ Utils.rawValueTimestampHeaders(value.priorValue(), timestamp),
+ Utils.rawValueTimestampHeaders(value.oldValue(), timestamp),
+ Utils.rawValueTimestampHeaders(value.newValue(), timestamp),
+ value.context()
+ );
+ }
+ // ValueTimestampHeaders -> plain would drop the per-value
headers/timestamps carried by a V4
+ // changelog. This only arises when a store that previously used a
headers store format is
+ // restored without one, i.e. a store-format downgrade. Downgrades are
not supported, so we
+ // fail fast here rather than silently discarding the stored header
data.
+ throw new StreamsException(
+ "Cannot restore suppress buffer '" + storeName + "': its changelog
contains header-aware "
+ + "(V4) records, but the store is configured without a headers
store format. "
+ + "Downgrading the store format is not supported."
+ );
+ }
+
+ private ValueTimestampHeadersSerializer<V>
valueTimestampHeadersSerializer() {
+ if (valueTimestampHeadersSerializer == null) {
+ valueTimestampHeadersSerializer = new
ValueTimestampHeadersSerializer<>(valueSerde.innerSerde().serializer());
+ }
+ return valueTimestampHeadersSerializer;
+ }
+
+ private ValueTimestampHeadersDeserializer<V>
valueTimestampHeadersDeserializer() {
+ if (valueTimestampHeadersDeserializer == null) {
+ valueTimestampHeadersDeserializer = new
ValueTimestampHeadersDeserializer<>(valueSerde.innerSerde().deserializer());
+ }
+ return valueTimestampHeadersDeserializer;
+ }
+
+ // Serialize a single value part. When storeHeaders is set, the value is
wrapped as a
+ // ValueTimestampHeaders blob carrying the given timestamp and headers;
otherwise the plain
+ // value bytes are stored (the pre-existing V3 behavior). Returns null for
a null value.
+ private byte[] serializeValuePart(final V value, final long timestamp,
final Headers headers) {
+ if (value == null) {
+ return null;
+ }
+ if (storeHeaders) {
+ return valueTimestampHeadersSerializer().serialize(changelogTopic,
ValueTimestampHeaders.make(value, timestamp, headers));
+ }
+ return valueSerde.innerSerde().serializer().serialize(changelogTopic,
headers, value);
+ }
+
+ // Deserialize a single stored value part into a ValueTimestampHeaders.
Only valid when
+ // storeHeaders is set (i.e. the part bytes are a ValueTimestampHeaders
blob).
+ private ValueTimestampHeaders<V> deserializeValuePart(final byte[] bytes) {
+ return bytes == null ? null :
valueTimestampHeadersDeserializer().deserialize(changelogTopic, bytes);
+ }
+
+ // Deserialize a single stored value part into the plain value, handling
both the headers-aware
+ // (ValueTimestampHeaders) and plain encodings.
+ private V deserializeValue(final byte[] bytes, final Headers
fallbackHeaders) {
+ if (bytes == null) {
+ return null;
+ }
+ if (storeHeaders) {
+ return
ValueTimestampHeaders.getValueOrNull(deserializeValuePart(bytes));
+ }
+ return
valueSerde.innerSerde().deserializer().deserialize(changelogTopic,
fallbackHeaders, bytes);
+ }
+
@Override
public boolean put(final long time,
final Record<K, Change<V>> record,
final ProcessorRecordContext recordContext) {
requireNonNull(record.value(), "value cannot be null");
requireNonNull(recordContext, "recordContext cannot be null");
- final Bytes serializedKey =
Bytes.wrap(keySerde.serializer().serialize(changelogTopic,
recordContext.headers(), record.key()));
- final Change<byte[]> serialChange =
valueSerde.serializeParts(changelogTopic, recordContext.headers(),
record.value());
-
+ // The record's own headers (not the processing context's) describe
the new value and must be
+ // the ones forwarded for this key on eviction.
+ final RecordHeaders headers = new RecordHeaders(record.headers());
+ final long timestamp = record.timestamp();
+ final Bytes serializedKey =
Bytes.wrap(keySerde.serializer().serialize(changelogTopic, headers,
record.key()));
final BufferValue buffered = getBuffered(serializedKey);
- final byte[] serializedPriorValue;
- if (buffered == null) {
- serializedPriorValue = serialChange.oldValue;
- } else {
- serializedPriorValue = buffered.priorValue();
+
+ // The context stored with the entry carries the currently-processed
record's headers; these
+ // are what get forwarded downstream when the (new) value is emitted.
Everything else is taken
+ // from the processing context, unchanged from before.
+ final ProcessorRecordContext effectiveContext = new
ProcessorRecordContext(
+ recordContext.timestamp(),
+ recordContext.offset(),
+ recordContext.partition(),
+ recordContext.topic(),
+ headers
+ );
+
+ // The old value's original headers/timestamp are not carried by the
incoming record. On an
+ // in-place update we recover them from the entry's previous new value
(whose value is exactly
+ // this update's old value); on the first insert for a key they are
genuinely unknown.
+ Headers oldHeaders = new RecordHeaders();
+ long oldTimestamp = RecordQueue.UNKNOWN;
+ if (storeHeaders && buffered != null) {
+ final ValueTimestampHeaders<V> previousNewValue =
deserializeValuePart(buffered.newValue());
+ if (previousNewValue != null) {
+ oldHeaders = previousNewValue.headers();
+ oldTimestamp = previousNewValue.timestamp();
+ }
}
+ final Change<V> change = record.value();
+ final byte[] newValue = serializeValuePart(change.newValue, timestamp,
headers);
+ // In plain mode the old value is still serialized with the current
record's headers, exactly
+ // as before, so the stored bytes are unchanged for non-header stores.
+ final byte[] oldValue = serializeValuePart(change.oldValue,
oldTimestamp, storeHeaders ? oldHeaders : headers);
Review Comment:
We actually need to serialize oldValue first (same as
`FullChangeSerde#serializeParts` which is called in the old code further above,
to ensure that tombstone are emitted with the correct header).
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
