@thetumbled Please add the following content to your PR description and select a checkbox:

- [ ] `doc` <!-- Your PR contains doc changes -->
- [ ] `doc-required` <!-- Your PR changes impact docs and you will update later -->
- [ ] `doc-not-needed` <!-- Your PR changes do not impact docs -->
- [ ] `doc-complete` <!-- Docs have been already added -->

This seems to be a bug in changes introduced by #14515 . @lordcheng10 Please review

There's a checkstyle failure in https://github.com/thetumbled/pulsar/actions/runs/3418835699/jobs/5691621649#step:8:996

But I'm still not sure about the following logic

                    // First try optimistic locking
                    long storedKey1 = table[bucket];
                    long storedKey2 = table[bucket + 1];
                    long storedValue1 = table[bucket + 2];
                    long storedValue2 = table[bucket + 3];

                    if (!acquiredLock && validate(stamp)) {
                        // The values we have read are consistent
                        if (key1 == storedKey1 && key2 == storedKey2) {
                            return new LongPair(storedValue1, storedValue2);

Even if storedKey1 and storeKey2 matched the expected keys, are storedValue1 and storedValue2 guaranteed to be exactly the associated values? This PR covered the case that table is shrinking, which makes the index invalid. However, could it happen that when retrieving table[bucket + 2] and table[bucket + 3], these values have been modified? @lhotari @lordcheng10 @thetumbled

There's a checkstyle failure in https://github.com/thetumbled/pulsar/actions/runs/3418835699/jobs/5691621649#step:8:996

done

But I'm still not sure about the following logic

                    // First try optimistic locking
                    long storedKey1 = table[bucket];
                    long storedKey2 = table[bucket + 1];
                    long storedValue1 = table[bucket + 2];
                    long storedValue2 = table[bucket + 3];

                    if (!acquiredLock && validate(stamp)) {
                        // The values we have read are consistent
                        if (key1 == storedKey1 && key2 == storedKey2) {
                            return new LongPair(storedValue1, storedValue2);

Even if storedKey1 and storeKey2 matched the expected keys, are storedValue1 and storedValue2 guaranteed to be exactly the associated values? This PR covered the case that table is shrinking, which makes the index invalid. However, could it happen that when retrieving table[bucket + 2] and table[bucket + 3], these values have been modified? @lhotari @lordcheng10 @thetumbled

if table[bucket + 2] and table[bucket + 3] is modified, there must be one thread having acquired write lock, which result into that validate(stamp) return false and fallback to pessimistic read.

Nice catch! And I think the ConcurrentLongLongPairHashMap.Section#forEach method may have the same problem. Could you help to confirm it?

And ConcurrentLongLongPairHashMap.Section#remove has the same issue. Please also handle it.

And ConcurrentLongLongPairHashMap.Section#remove has the same issue. Please also handle it.

i don't find that ConcurrentLongLongPairHashMap.Section#remove do use any optimistic read lock but write lock.

Yes, only tryOptimisticRead may cause this issue.

Oh my fault, here is a writeLock in the ConcurrentLongLongPairHashMap.Section#remove.

Nice catch! And I think the ConcurrentLongLongPairHashMap.Section#forEach method may have the same problem. Could you help to confirm it?

i find that, even if there is any thread acquired write lock and trigger shrink/expand process, the local variable table will not be updated, so the index used in loop will not exceed the bound.

The ConcurrentLongLongPairHashMap comes from https://github.com/apache/bookkeeper/blob/master/bookkeeper-server/src/main/java/org/apache/bookkeeper/util/collections/ConcurrentLongLongHashMap.java

Maybe there is the same bug in bookkeeper, or it only occur in getting stored value?

the local variable table will not be updated.

@thetumbled It's not true. table variable is a reference not a copy, so its length may change.

the local variable table will not be updated.

@thetumbled It's not true. table variable is a reference not a copy, so its length may change.

In the rehash method, the member this.table of Section will be reallocated a new long[] array, which is different from old table member.

        private void rehash(int newCapacity) {
            long[] newTable = new long[4 * newCapacity];
            ...
            table = newTable;
            ...
        }

For example:

    public static void main(String[] args) {
        int[] table = new int[]{1, 2};
        int[] localTable = table;
        System.out.println(localTable.length);
        // rehash table to new array.
        table = new int[]{3};
        System.out.println(localTable.length);
    }

the result will be following:

Oh I see, thanks for your explanation.

The ConcurrentLongLongPairHashMap comes from https://github.com/apache/bookkeeper/blob/master/bookkeeper-server/src/main/java/org/apache/bookkeeper/util/collections/ConcurrentLongLongHashMap.java

Maybe there is the same bug in bookkeeper, or it only occur in getting stored value?

it seems that there is the same bug in bookkeeper.

In the rehash method, the member this.table of Section will be reallocated a new long[] array, which is different from old table member.

        private void rehash(int newCapacity) {
            long[] newTable = new long[4 * newCapacity];
            ...
            table = newTable;
            ...
        }

This inspires me. Can we fix the get issue in the same way?

            boolean acquiredLock = false;
            // add local variable here, so OutOfBound won't happen
            long[] table = this.table;
            // calculate table.length / 4 as capacity to avoid rehash changing capacity
            int bucket = signSafeMod(keyHash, table.length / 4);

            try {
                while (true) {
                    long storedKey1 = table[bucket];
                    long storedKey2 = table[bucket + 1];
                    long storedValue1 = table[bucket + 2];
                    long storedValue2 = table[bucket + 3];
                    
                    // compare table to ensure not rehash
                    if (!acquiredLock && validate(stamp) && table == this.table) {
                        if (key1 == storedKey1 && key2 == storedKey2) {
                            return new LongPair(storedValue1, storedValue2);
                        } else if (storedKey1 == EmptyKey) {
                            return null;
                        }
                    } else {
                        if (!acquiredLock) {
                            stamp = readLock();
                            acquiredLock = true;
                            // update local variable
                            table = this.table;
                            bucket = signSafeMod(keyHash, capacity);
                            storedKey1 = table[bucket];
                            storedKey2 = table[bucket + 1];
                            storedValue1 = table[bucket + 2];
                            storedValue2 = table[bucket + 3];
                        }

/pulsarbot run-failure-checks

/pulsarbot run-failure-checks