[
https://issues.apache.org/jira/browse/FLINK-32701?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17751083#comment-17751083
]
Puneet Duggal commented on FLINK-32701:
---------------------------------------
[~martijnvisser] [~Juntao Hu] [~nicholasjiang] During the execution of a job,
two primary sources of memory leak have been identified in the CEP Operator:
# {{NFAState}}
# {{{}SharedBuffer.eventsCount{}}}.
Implementations to resolve these memory leaks are as follows:
*NFAState Leak Resolution*
The {{NFAState}} keyed state in CEP Operator contains two states:
* {{Queue<ComputationState> partialMatches}}
* {{Queue<ComputationState> completedMatches}}
Despite all events for a key being processed (either matches or timed out), the
{{partialMatches}} still retains the starting state for that key. This occurs
for every key encountered by CEP throughout the job execution, leading to a
memory leak. To mitigate this, a check has been introduced: once all matches
have been processed and the time for all states advances based on the
watermark, the {{NFAState}} is cleared if {{completedMatches}} is empty and
{{partialMatches}} only contains a single state (the starting state).
{code:java}
// STEP 4
updateNFA(nfaState);
// In order to remove dangling partial matches
if (nfaState.getPartialMatches().size() == 1 &&
nfaState.getCompletedMatches().isEmpty()) {
computationStates.clear();
}
{code}
The applied fix has been tested with the existing set of Flink unit test cases,
all of which have passed. The fix has also been verified against our specific
use case scenarios, and it functions as expected.
*SharedBuffer.EventsCount Leak Resolution*
The {{eventsCount}} in the shared buffer is responsible for maintaining the
mapping of timestamp and {{eventId}} for each event for a key. As the watermark
surpasses the timestamp of an event, CEP continues to remove mappings from
{{{}eventsCount{}}}. However, an empty map state for a key still consumes
memory, resulting in a memory leak. To rectify this, a check has been added: if
the {{eventsCount}} map state is empty after the CEP Operator advances time
(removing events and matches with a timestamp earlier than the watermark), it
is cleared.
This fix, upon testing, resulted in the failure of two unit test cases. These
failures occurred because the tests assert a fixed number of total state writes
in the CEP Operator when evaluating a pattern sequence. As expected, this
number has increased because we are clearing the {{eventsCount}} map. However,
when tested against our specific use case scenarios, the fix functioned
correctly.
{code:java}
void advanceTime(long timestamp) throws Exception {
Iterator<Long> iterator = eventsCount.keys().iterator();
while (iterator.hasNext()) {
Long next = iterator.next();
if (next < timestamp) {
iterator.remove();
}
}
//memory leak resolution
if (eventsCount.isEmpty()) {
eventsCount.clear();
}
}
{code}
Please let me know if there are any concerns or questions.
> Potential Memory Leak in Flink CEP due to Persistent Starting States in
> NFAState
> --------------------------------------------------------------------------------
>
> Key: FLINK-32701
> URL: https://issues.apache.org/jira/browse/FLINK-32701
> Project: Flink
> Issue Type: Bug
> Components: Library / CEP
> Affects Versions: 1.17.0, 1.17.1
> Reporter: Puneet Duggal
> Priority: Critical
> Attachments: Screenshot 2023-07-26 at 11.45.06 AM.png, Screenshot
> 2023-07-26 at 11.50.28 AM.png
>
>
> Our team has encountered a potential memory leak issue while working with the
> Complex Event Processing (CEP) library in Flink v1.17.
> h2. Context
> The CEP Operator maintains a keyed state called NFAState, which holds two
> queues: one for partial matches and one for completed matches. When a key is
> first encountered, the CEP creates a starting computation state and stores it
> in the partial matches queue. As more events occur that match the defined
> conditions (e.g., a TAKE condition), additional computation states get added
> to the queue, with their specific type (normal, pending, end) depending on
> the pattern sequence.
> However, I have noticed that the starting computation state remains in the
> partial matches queue even after the pattern sequence has been completely
> matched. This is also the case for keys that have already timed out. As a
> result, the state gets stored for all keys that the CEP ever encounters,
> leading to a continual increase in the checkpoint size.
> h2. How to reproduce this
> # Pattern Sequence - A not_followed_by B within 5 mins
> # Time Characteristic - EventTime
> # StateBackend - FsStateBackend
> On my local machine, I started this pipeline and started sending events at
> the rate of 10 events per second (only A) and as expected after 5 mins, CEP
> started sending pattern matched output with the same rate. But the issue was
> that after every 2 mins (checkpoint interval), checkpoint size kept on
> increasing. Expectation was that after 5 mins (2-3 checkpoints), checkpoint
> size will remain constant since any window of 5 mins will consist of the same
> number of unique keys (older ones will get matched or timed out hence removed
> from state). But as you can see below attached images, checkpoint size kept
> on increasing till 40 checkpoints (around 1.5hrs).
> P.S. - After 3 checkpoints (6 mins), the checkpoint size was around 1.78MB.
> Hence assumption is that ideal checkpoint size for a 5 min window should be
> less than 1.78MB.
> As you can see after 39 checkpoints, I triggered a savepoint for this
> pipeline. After that I used a savepoint reader to investigate what all is
> getting stored in CEP states. Below code investigates NFAState of CEPOperator
> for potential memory leak.
> {code:java}
> import lombok.AllArgsConstructor;
> import lombok.Data;
> import lombok.NoArgsConstructor;
> import org.apache.flink.api.common.state.ValueState;
> import org.apache.flink.api.common.state.ValueStateDescriptor;
> import org.apache.flink.cep.nfa.NFAState;
> import org.apache.flink.cep.nfa.NFAStateSerializer;
> import org.apache.flink.configuration.Configuration;
> import org.apache.flink.runtime.state.filesystem.FsStateBackend;
> import org.apache.flink.state.api.OperatorIdentifier;
> import org.apache.flink.state.api.SavepointReader;
> import org.apache.flink.state.api.functions.KeyedStateReaderFunction;
> import org.apache.flink.streaming.api.datastream.DataStream;
> import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
> import org.apache.flink.util.Collector;
> import org.junit.jupiter.api.Test;
> import java.io.Serializable;
> import java.util.Objects;
> public class NFAStateReaderTest {
> private static final String NFA_STATE_NAME = "nfaStateName";
> @Test
> public void testNfaStateReader() throws Exception {
> StreamExecutionEnvironment environment =
> StreamExecutionEnvironment.getExecutionEnvironment();
> SavepointReader savepointReader =
> SavepointReader.read(environment,
> "file:///opt/flink/savepoints/savepoint-093404-9bc0a38654df", new
> FsStateBackend("file:///abc"));
> DataStream<NFAStateOutput> stream =
> savepointReader.readKeyedState(OperatorIdentifier.forUid("select_pattern_events"),
> new NFAStateReaderTest.NFAStateReaderFunction());
> stream.print();
> environment.execute();
> }
> static class NFAStateReaderFunction extends
> KeyedStateReaderFunction<DynamicTuple, NFAStateOutput> {
> private ValueState<NFAState> computationStates;
> private static Long danglingNfaCount = 0L;
> private static Long newNfaCount = 0L;
> private static Long minTimestamp = Long.MAX_VALUE;
> private static Long minKeyForCurrentNfa = Long.MAX_VALUE;
> private static Long minKeyForDanglingNfa = Long.MAX_VALUE;
> private static Long maxKeyForDanglingNfa = Long.MIN_VALUE;
> private static Long maxKeyForCurrentNfa = Long.MIN_VALUE;
> @Override
> public void open(Configuration parameters) {
> computationStates = getRuntimeContext().getState(new
> ValueStateDescriptor<>(NFA_STATE_NAME, new NFAStateSerializer()));
> }
> @Override
> public void readKey(DynamicTuple key, Context ctx,
> Collector<NFAStateOutput> out) throws Exception {
> NFAState nfaState = computationStates.value();
> if
> (Objects.requireNonNull(nfaState.getPartialMatches().peek()).getStartTimestamp()
> != -1) {
> minTimestamp = Math.min(minTimestamp,
> nfaState.getPartialMatches().peek().getStartTimestamp());
> minKeyForCurrentNfa = Math.min(minKeyForCurrentNfa,
> Long.parseLong(key.getTuple().getField(0)));
> maxKeyForCurrentNfa = Math.max(maxKeyForCurrentNfa,
> Long.parseLong(key.getTuple().getField(0)));
> newNfaCount++;
> } else {
> danglingNfaCount++;
> minKeyForDanglingNfa = Math.min(minKeyForDanglingNfa,
> Long.parseLong(key.getTuple().getField(0)));
> maxKeyForDanglingNfa = Math.max(maxKeyForDanglingNfa,
> Long.parseLong(key.getTuple().getField(0)));
> }
> NFAStateOutput nfaStateOutput =
> new NFAStateOutput(
> danglingNfaCount,
> minTimestamp,
> newNfaCount,
> minKeyForCurrentNfa,
> maxKeyForCurrentNfa,
> minKeyForDanglingNfa,
> maxKeyForDanglingNfa);
> out.collect(nfaStateOutput);
> }
> }
> @Data
> @NoArgsConstructor
> @AllArgsConstructor
> static class NFAStateOutput implements Serializable {
> private Long danglingNfaCount;
> private Long minTimestamp;
> private Long newNfaCount;
> private Long minKeyForCurrentNfa;
> private Long maxKeyForCurrentNfa;
> private Long minKeyForDanglingNfa;
> private Long maxKeyForDanglingNfa;
> }
> }
> {code}
>
> As an output it printed nfaStateOutput for each key but since all the
> attributes in nfaStateOutput are aggregates, hence finalOutput printed was
> {code:java}
> NFAStateReaderTest.NFAStateOutput(danglingNfaCount=34391,
> minTimestamp=1690359951958, newNfaCount=3000, minKeyForCurrentNfa=6244230,
> maxKeyForCurrentNfa=6247229, minKeyForDanglingNfa=629818,
> maxKeyForDanglingNfa=6244229){code}
>
> As we can see, checkpoint is storing approximately 34391 dangling states (for
> keys which have expired (matched or timed out) ) whereas there are only 3000
> active keys (for which there are partial matches which are eligible for
> further pattern sequence matching) which is expected since throughput is 10
> events per second which amounts to 3000 unique keys in 5 mins.
> h2. Questions
> Hence, I am curious about the reasoning behind this design choice,
> specifically why the starting state remains in the partial matches queue for
> all keys, even those that have either timed out or completed their matches.
> Additionally, I am wondering what the implications would be if we were to
> delete this starting state assuming that
> # it is the only state left in the partial match queue.
> # The completed match queue in nfaState is empty.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
