xiangfu0 commented on code in PR #19539:
URL: https://github.com/apache/pinot/pull/19539#discussion_r4067648333
##########
pinot-broker/src/main/java/org/apache/pinot/broker/routing/segmentpruner/SinglePartitionColumnSegmentPruner.java:
##########
@@ -126,30 +131,180 @@ private boolean isPartitionMatch(Expression
filterExpression, SegmentPartitionIn
return false;
case EQUALS: {
Identifier identifier = operands.get(0).getIdentifier();
- if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- return
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
- } else {
- return true;
- }
+ return identifier == null ||
!identifier.getName().equals(_partitionColumn)
+ ||
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
}
case IN: {
Identifier identifier = operands.get(0).getIdentifier();
+ if (identifier == null ||
!identifier.getName().equals(_partitionColumn)) {
+ return true;
+ }
+ for (int i = 1; i < operands.size(); i++) {
+ if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ return true;
+ }
+ }
+ return false;
+ }
+ default:
+ return true;
+ }
+ }
+
+ /// All prepared predicates and hashes belong to one prune call; refreshes
and other queries share none of this state.
+ private final class QueryPartitionMatcher {
+ private final Expression _filterExpression;
+ private final Map<PartitionFunctionKey, PreparedPredicate> _predicates =
new HashMap<>();
+ private final PartitionFunctionLookup _lookup = new
PartitionFunctionLookup();
+ private PreparedPredicate _lastPredicate;
+
+ private QueryPartitionMatcher(Expression filterExpression) {
+ _filterExpression = filterExpression;
+ }
+
+ private boolean matches(SegmentPartitionInfo partitionInfo) {
+ // Segment metadata contains distinct function instances. Avoid
allocating a key per segment for the common
+ // case where those instances have identical configuration.
+ if (_lastPredicate == null || !_lookup.matches(partitionInfo)) {
+ // This reusable lookup probe is never inserted. Only a previously
unseen configuration allocates a stored key.
+ _lookup._partitionInfo = partitionInfo;
+ _lastPredicate = _predicates.get(_lookup);
+ if (_lastPredicate == null) {
+ _lastPredicate = new PreparedPredicate(_filterExpression,
partitionInfo.getPartitionFunction());
+ _predicates.put(new CachedPartitionFunctionKey(partitionInfo),
_lastPredicate);
+ }
+ }
+ return _lastPredicate.matches(partitionInfo.getPartitions());
+ }
+ }
+
+ /// Interprets each visited predicate once and hashes IN values only as far
as short-circuit evaluation requires.
+ private final class PreparedPredicate {
+ private final Expression _expression;
+ private final PartitionFunction _partitionFunction;
+ private FilterKind _filterKind;
+ private List<Expression> _operands;
+ private PreparedPredicate[] _children;
+ private Integer[] _partitionIds;
+
+ private PreparedPredicate(Expression expression, PartitionFunction
partitionFunction) {
+ _expression = expression;
+ _partitionFunction = partitionFunction;
+ }
+
+ private void prepare() {
+ Function function = _expression.getFunctionCall();
+ _filterKind = FilterKind.valueOf(function.getOperator());
+ _operands = function.getOperands();
+ if (_filterKind == FilterKind.AND || _filterKind == FilterKind.OR) {
+ _children = new PreparedPredicate[_operands.size()];
+ for (int i = 0; i < _children.length; i++) {
+ _children[i] = new PreparedPredicate(_operands.get(i),
_partitionFunction);
+ }
+ } else if (_filterKind == FilterKind.EQUALS || _filterKind ==
FilterKind.IN) {
+ Identifier identifier = _operands.get(0).getIdentifier();
if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- int numOperands = operands.size();
- for (int i = 1; i < numOperands; i++) {
- if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ _partitionIds = new Integer[_filterKind == FilterKind.EQUALS ? 1 :
_operands.size() - 1];
+ }
+ }
+ }
+
+ private boolean matches(Set<Integer> partitions) {
+ if (_filterKind == null) {
+ prepare();
+ }
+ switch (_filterKind) {
+ case AND:
+ for (PreparedPredicate child : _children) {
+ if (!child.matches(partitions)) {
+ return false;
+ }
+ }
+ return true;
+ case OR:
+ for (PreparedPredicate child : _children) {
+ if (child.matches(partitions)) {
+ return true;
+ }
+ }
+ return false;
+ case EQUALS:
+ case IN:
+ if (_partitionIds == null) {
+ return true;
+ }
+ for (int i = 0; i < _partitionIds.length; i++) {
+ Integer partitionId = _partitionIds[i];
+ if (partitionId == null) {
+ partitionId =
_partitionFunction.getPartition(RequestContextUtils.getStringValue(_operands.get(i
+ 1)));
+ _partitionIds[i] = partitionId;
+ }
+ if (partitions.contains(partitionId)) {
return true;
}
}
return false;
- } else {
+ default:
return true;
- }
}
- default:
- return true;
+ }
+ }
+
+ /// Uses all recorded constructor inputs to identify equivalent partition
functions.
+ private abstract static class PartitionFunctionKey {
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): The wrapper/key
and its map have been removed. Compatibility now belongs to PartitionFunction;
metadata objects remain unchanged and the cache is local to a single prune
call. This addresses the wrapper concern without expanding SegmentPartitionInfo.
##########
pinot-broker/src/main/java/org/apache/pinot/broker/routing/segmentpruner/SinglePartitionColumnSegmentPruner.java:
##########
@@ -95,10 +98,12 @@ public Set<String> prune(BrokerRequest brokerRequest,
Set<String> segments) {
return segments;
}
Set<String> selectedSegments = new HashSet<>();
+ // A singleton has no repeated work to reuse. Keep its evaluation free of
predicate/cache setup.
+ QueryPartitionMatcher matcher = segments.size() > 1 ? new
QueryPartitionMatcher(filterExpression) : null;
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): The current design
deliberately retains the original evaluator below 256 candidates to avoid
preparation overhead. The prepared path is lazy, and the expanded shared
fixtures now exercise 255/256 candidates plus a forced uncached path through
enablePartitionPruningCache=false. This addresses the coverage and operational
fallback concerns, but not the preference for maintaining a single evaluator.
Leaving that design choice open rather than claiming it is resolved.
##########
pinot-broker/src/main/java/org/apache/pinot/broker/routing/segmentpruner/SinglePartitionColumnSegmentPruner.java:
##########
@@ -126,30 +131,180 @@ private boolean isPartitionMatch(Expression
filterExpression, SegmentPartitionIn
return false;
case EQUALS: {
Identifier identifier = operands.get(0).getIdentifier();
- if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- return
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
- } else {
- return true;
- }
+ return identifier == null ||
!identifier.getName().equals(_partitionColumn)
+ ||
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
}
case IN: {
Identifier identifier = operands.get(0).getIdentifier();
+ if (identifier == null ||
!identifier.getName().equals(_partitionColumn)) {
+ return true;
+ }
+ for (int i = 1; i < operands.size(); i++) {
+ if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ return true;
+ }
+ }
+ return false;
+ }
+ default:
+ return true;
+ }
+ }
+
+ /// All prepared predicates and hashes belong to one prune call; refreshes
and other queries share none of this state.
+ private final class QueryPartitionMatcher {
+ private final Expression _filterExpression;
+ private final Map<PartitionFunctionKey, PreparedPredicate> _predicates =
new HashMap<>();
+ private final PartitionFunctionLookup _lookup = new
PartitionFunctionLookup();
+ private PreparedPredicate _lastPredicate;
+
+ private QueryPartitionMatcher(Expression filterExpression) {
+ _filterExpression = filterExpression;
+ }
+
+ private boolean matches(SegmentPartitionInfo partitionInfo) {
+ // Segment metadata contains distinct function instances. Avoid
allocating a key per segment for the common
+ // case where those instances have identical configuration.
+ if (_lastPredicate == null || !_lookup.matches(partitionInfo)) {
+ // This reusable lookup probe is never inserted. Only a previously
unseen configuration allocates a stored key.
+ _lookup._partitionInfo = partitionInfo;
+ _lastPredicate = _predicates.get(_lookup);
+ if (_lastPredicate == null) {
+ _lastPredicate = new PreparedPredicate(_filterExpression,
partitionInfo.getPartitionFunction());
+ _predicates.put(new CachedPartitionFunctionKey(partitionInfo),
_lastPredicate);
+ }
+ }
+ return _lastPredicate.matches(partitionInfo.getPartitions());
+ }
+ }
+
+ /// Interprets each visited predicate once and hashes IN values only as far
as short-circuit evaluation requires.
+ private final class PreparedPredicate {
+ private final Expression _expression;
+ private final PartitionFunction _partitionFunction;
+ private FilterKind _filterKind;
+ private List<Expression> _operands;
+ private PreparedPredicate[] _children;
+ private Integer[] _partitionIds;
+
+ private PreparedPredicate(Expression expression, PartitionFunction
partitionFunction) {
+ _expression = expression;
+ _partitionFunction = partitionFunction;
+ }
+
+ private void prepare() {
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): Keeping decoding
lazy is necessary for the current evaluator's short-circuit behavior: an
invalid operator or literal in an unvisited AND/OR/IN branch must not start
throwing. The shared fixtures exercise this at 255 and 256 candidates and with
caching disabled. Constructor preparation of the whole tree would violate those
cases; the lazy fields and their purpose are retained.
##########
pinot-broker/src/main/java/org/apache/pinot/broker/routing/segmentpruner/SinglePartitionColumnSegmentPruner.java:
##########
@@ -126,30 +131,180 @@ private boolean isPartitionMatch(Expression
filterExpression, SegmentPartitionIn
return false;
case EQUALS: {
Identifier identifier = operands.get(0).getIdentifier();
- if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- return
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
- } else {
- return true;
- }
+ return identifier == null ||
!identifier.getName().equals(_partitionColumn)
+ ||
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
}
case IN: {
Identifier identifier = operands.get(0).getIdentifier();
+ if (identifier == null ||
!identifier.getName().equals(_partitionColumn)) {
+ return true;
+ }
+ for (int i = 1; i < operands.size(); i++) {
+ if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ return true;
+ }
+ }
+ return false;
+ }
+ default:
+ return true;
+ }
+ }
+
+ /// All prepared predicates and hashes belong to one prune call; refreshes
and other queries share none of this state.
+ private final class QueryPartitionMatcher {
+ private final Expression _filterExpression;
+ private final Map<PartitionFunctionKey, PreparedPredicate> _predicates =
new HashMap<>();
+ private final PartitionFunctionLookup _lookup = new
PartitionFunctionLookup();
+ private PreparedPredicate _lastPredicate;
+
+ private QueryPartitionMatcher(Expression filterExpression) {
+ _filterExpression = filterExpression;
+ }
+
+ private boolean matches(SegmentPartitionInfo partitionInfo) {
+ // Segment metadata contains distinct function instances. Avoid
allocating a key per segment for the common
+ // case where those instances have identical configuration.
+ if (_lastPredicate == null || !_lookup.matches(partitionInfo)) {
+ // This reusable lookup probe is never inserted. Only a previously
unseen configuration allocates a stored key.
+ _lookup._partitionInfo = partitionInfo;
+ _lastPredicate = _predicates.get(_lookup);
+ if (_lastPredicate == null) {
+ _lastPredicate = new PreparedPredicate(_filterExpression,
partitionInfo.getPartitionFunction());
+ _predicates.put(new CachedPartitionFunctionKey(partitionInfo),
_lastPredicate);
+ }
+ }
+ return _lastPredicate.matches(partitionInfo.getPartitions());
+ }
+ }
+
+ /// Interprets each visited predicate once and hashes IN values only as far
as short-circuit evaluation requires.
+ private final class PreparedPredicate {
+ private final Expression _expression;
+ private final PartitionFunction _partitionFunction;
+ private FilterKind _filterKind;
+ private List<Expression> _operands;
+ private PreparedPredicate[] _children;
+ private Integer[] _partitionIds;
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): Kept the ordered
visited-prefix list intentionally. Deduplicating eagerly into an IntSet would
evaluate remaining literals after an earlier match, changing lazy-error
behavior and adding work for early-matching long IN filters. The cache computes
each visited literal once, while incompatible functions neither read nor extend
it. Deduplication can be a separate measured optimization without complicating
this query-local change.
##########
pinot-broker/src/main/java/org/apache/pinot/broker/routing/segmentpruner/SinglePartitionColumnSegmentPruner.java:
##########
@@ -126,30 +131,180 @@ private boolean isPartitionMatch(Expression
filterExpression, SegmentPartitionIn
return false;
case EQUALS: {
Identifier identifier = operands.get(0).getIdentifier();
- if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- return
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
- } else {
- return true;
- }
+ return identifier == null ||
!identifier.getName().equals(_partitionColumn)
+ ||
partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(1))));
}
case IN: {
Identifier identifier = operands.get(0).getIdentifier();
+ if (identifier == null ||
!identifier.getName().equals(_partitionColumn)) {
+ return true;
+ }
+ for (int i = 1; i < operands.size(); i++) {
+ if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
+
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ return true;
+ }
+ }
+ return false;
+ }
+ default:
+ return true;
+ }
+ }
+
+ /// All prepared predicates and hashes belong to one prune call; refreshes
and other queries share none of this state.
+ private final class QueryPartitionMatcher {
+ private final Expression _filterExpression;
+ private final Map<PartitionFunctionKey, PreparedPredicate> _predicates =
new HashMap<>();
+ private final PartitionFunctionLookup _lookup = new
PartitionFunctionLookup();
+ private PreparedPredicate _lastPredicate;
+
+ private QueryPartitionMatcher(Expression filterExpression) {
+ _filterExpression = filterExpression;
+ }
+
+ private boolean matches(SegmentPartitionInfo partitionInfo) {
+ // Segment metadata contains distinct function instances. Avoid
allocating a key per segment for the common
+ // case where those instances have identical configuration.
+ if (_lastPredicate == null || !_lookup.matches(partitionInfo)) {
+ // This reusable lookup probe is never inserted. Only a previously
unseen configuration allocates a stored key.
+ _lookup._partitionInfo = partitionInfo;
+ _lastPredicate = _predicates.get(_lookup);
+ if (_lastPredicate == null) {
+ _lastPredicate = new PreparedPredicate(_filterExpression,
partitionInfo.getPartitionFunction());
+ _predicates.put(new CachedPartitionFunctionKey(partitionInfo),
_lastPredicate);
+ }
+ }
+ return _lastPredicate.matches(partitionInfo.getPartitions());
+ }
+ }
+
+ /// Interprets each visited predicate once and hashes IN values only as far
as short-circuit evaluation requires.
+ private final class PreparedPredicate {
+ private final Expression _expression;
+ private final PartitionFunction _partitionFunction;
+ private FilterKind _filterKind;
+ private List<Expression> _operands;
+ private PreparedPredicate[] _children;
+ private Integer[] _partitionIds;
+
+ private PreparedPredicate(Expression expression, PartitionFunction
partitionFunction) {
+ _expression = expression;
+ _partitionFunction = partitionFunction;
+ }
+
+ private void prepare() {
+ Function function = _expression.getFunctionCall();
+ _filterKind = FilterKind.valueOf(function.getOperator());
+ _operands = function.getOperands();
+ if (_filterKind == FilterKind.AND || _filterKind == FilterKind.OR) {
+ _children = new PreparedPredicate[_operands.size()];
+ for (int i = 0; i < _children.length; i++) {
+ _children[i] = new PreparedPredicate(_operands.get(i),
_partitionFunction);
+ }
+ } else if (_filterKind == FilterKind.EQUALS || _filterKind ==
FilterKind.IN) {
+ Identifier identifier = _operands.get(0).getIdentifier();
if (identifier != null &&
identifier.getName().equals(_partitionColumn)) {
- int numOperands = operands.size();
- for (int i = 1; i < numOperands; i++) {
- if
(partitionInfo.getPartitions().contains(partitionInfo.getPartitionFunction()
-
.getPartition(RequestContextUtils.getStringValue(operands.get(i))))) {
+ _partitionIds = new Integer[_filterKind == FilterKind.EQUALS ? 1 :
_operands.size() - 1];
+ }
+ }
+ }
+
+ private boolean matches(Set<Integer> partitions) {
+ if (_filterKind == null) {
+ prepare();
+ }
+ switch (_filterKind) {
+ case AND:
+ for (PreparedPredicate child : _children) {
+ if (!child.matches(partitions)) {
+ return false;
+ }
+ }
+ return true;
+ case OR:
+ for (PreparedPredicate child : _children) {
+ if (child.matches(partitions)) {
+ return true;
+ }
+ }
+ return false;
+ case EQUALS:
+ case IN:
+ if (_partitionIds == null) {
+ return true;
+ }
+ for (int i = 0; i < _partitionIds.length; i++) {
+ Integer partitionId = _partitionIds[i];
+ if (partitionId == null) {
+ partitionId =
_partitionFunction.getPartition(RequestContextUtils.getStringValue(_operands.get(i
+ 1)));
+ _partitionIds[i] = partitionId;
+ }
+ if (partitions.contains(partitionId)) {
return true;
}
}
return false;
- } else {
+ default:
return true;
- }
}
- default:
- return true;
+ }
+ }
+
+ /// Uses all recorded constructor inputs to identify equivalent partition
functions.
+ private abstract static class PartitionFunctionKey {
+ abstract SegmentPartitionInfo getPartitionInfo();
+
+ final boolean matches(SegmentPartitionInfo partitionInfo) {
+ SegmentPartitionInfo current = getPartitionInfo();
+ PartitionFunction function = current.getPartitionFunction();
+ PartitionFunction otherFunction = partitionInfo.getPartitionFunction();
+ return function.getClass() == otherFunction.getClass() &&
function.getName().equals(otherFunction.getName())
+ && function.getNumPartitions() == otherFunction.getNumPartitions()
+ && function.getPartitionIdNormalizer() ==
otherFunction.getPartitionIdNormalizer()
+ && Objects.equals(current.getPartitionFunctionConfig(),
partitionInfo.getPartitionFunctionConfig());
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): The full
config-map comparison has been removed. The interface default checks only
config emptiness (or instance identity) and small function attributes;
BoundedColumnValue's independently deserialized large configs conservatively
fall back. The large-config/unrelated-column fixture now reaches 255/256
candidates and the cache-disabled path. No metadata interning or per-segment
scan of the 100 KB string is introduced.
##########
pinot-segment-spi/src/main/java/org/apache/pinot/segment/spi/partition/PartitionFunction.java:
##########
@@ -63,6 +63,26 @@ default Map<String, String> getFunctionConfig() {
return null;
}
+ /// Returns whether the exposed function configuration is null or empty.
This does not imply that other settings,
+ /// such as the partition id normalizer, have their default values.
+ @JsonIgnore
+ default boolean hasEmptyConfig() {
+ Map<String, String> config = getFunctionConfig();
+ return config == null || config.isEmpty();
+ }
+
+ /// Returns whether partition ids computed by this function can be reused
for the non-null `other` function.
+ /// A true result must guarantee identical results for every input value.
False is conservative, not proof that the
+ /// functions differ. Implementations with additional output-affecting state
must account for it in this method.
+ ///
+ /// The default only permits matching functions with empty exposed
configurations, without comparing config contents.
+ /// Configured implementations may override this method to compare their
effective settings.
+ default boolean canReusePartitionIds(PartitionFunction other) {
Review Comment:
Follow-up in d9135a1518 (38 focused Apache cases passed): Clarified the SPI
contract explicitly: reflexive, symmetric, transitive, and equal partition IDs
for every input; coarser/superset relations are not sufficient. Added the
same-instance fast path, including configured instances, and extended the
compatibility test for reflexivity and a three-instance compatible group.
Callers may safely populate the cache from any compatible member.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
