...

Repository queries taking collection parameters are limited to 32K collection items.
For example, given a FragmentRepository method:

@Query(value = "SELECT * FROM fragment WHERE anchor_id = :anchorId AND xpath IN :xpaths", nativeQuery = true)
List<FragmentEntity> findAllByAnchorIdAndXpathIn(int anchorId, Collection<String> xpaths);

...

Once the solution is implemented, these tests will be updated to show that no exception is thrown, with even more items e.g. 40,000.

Proposal #1:

...

Batching at 32,000

This solution adds a wrapper method that partitions the input collection into lists of 32,000 and calls the original repository method on each batch:

@Query(value = "SELECT * FROM fragment WHERE anchor_id = :anchorId AND xpath IN :xpaths", nativeQuery = true)
List<FragmentEntity> findAllByAnchorIdAndXpathIn(int anchorId, Collection<String> xpaths);

default List<FragmentEntity> findAllByAnchorIdAndXpathInBatch(final int anchorId, final List<String> xpaths) {
    List<FragmentEntity> results = new ArrayList<>(xpaths.size());
    for (final Collection<String> xpathsBatch : Lists.partition(xpaths, 32000)) {
        results.addAll(findAllByAnchorIdAndXpathIn(anchorId, xpathsBatch));
    }
    return results;
}

...

This solution is largely self-documenting, as the 32K limit is clearly seen in the code.

This solution still involves having a separate query parameter for each collection item - we are avoiding the limit by simply splitting into multiple SQL queries:

@Query(value = "SELECT * FROM fragment WHERE anchor_id = :anchorId? AND xpath IN :xpaths", nativeQuery = true)
List<FragmentEntity> findAllByAnchorIdAndXpathIn(int anchorId, Collection<String> xpaths);

default List<FragmentEntity> findAllByAnchorIdAndXpathInBatch(final int anchorId, final List<String> xpaths) {
    if (xpaths.size() < 32000) {
        return findAllByAnchorIdAndXpathIn(anchorId, xpathsBatch);
    }
    List<FragmentEntity> results = new ArrayList<>(xpaths.size());
    for (final Collection<String> xpathsBatch : Lists.partition(xpaths, 32000)) {
        results.addAll(findAllByAnchorIdAndXpathIn(anchorId, xpathsBatch));
    }
    return results;
}(?,?,?,?,?,?...,?)
SELECT * FROM fragment WHERE anchor_id = ? AND xpath IN (?,?,?,?,?,?...,?)
SELECT * FROM fragment WHERE anchor_id = ? AND xpath IN (?,?,?,?)

Observations

Performance testing is inconclusive (so far)

Testing with a batch size of 1000 (not 32000) showed a performance drop of 30% at 20000 items. More testing is needed to see if this applies with batch size of 32000, or if the test involved bad luck.

Advantage: Solution is easy to understand

This code is largely self-documenting, as the 32K limit is clearly seen in the code.

Batching on repository methods versus API methods

This solution should work equally well if batching at the public API methods instead of each repository method.

Proposal #2: Postgres arrays

This solution uses a third-party library to add PostGres Postgres array support to queries. Postgres arrays

Adding the dependency to cps-ri/pom.xml:

...

This solution involves creating two versions of each repository method taking a collection parameter - one that uses an taking a Java array, and one using a Collection:

@Query(value = "SELECT * FROM fragment WHERE anchor_id = :anchorId AND xpath = ANY(:xpaths)", nativeQuery = true)
List<FragmentEntity> findAllByAnchorIdAndXpathIn(@Param("anchorId") int anchorId, @Param("xpaths") String[] xpaths);

default List<FragmentEntity> findAllByAnchorIdAndXpathIn(final int anchorId, final Collection<String> xpaths) {
    return findAllByAnchorIdAndXpathIn(anchorId, xpaths.toArray(new String[0]));
}

This would result in the following a single SQL query, given any number of xpaths:

...

As seen, only two SQL query parameters will be ever used. This is because the array is binary-encoded and sent as a single parameter.

Using Postgres arrays can simplify existing code

Testing so far has shown no noticeable performance difference with master branch.

Observations

Performance is unchanged

Testing shows there was no increase or decrease in performance compared to master branch after arrays were used.

Disadvantage: extra dependency

Extra dependencies could add possible risks of dependency issues, security issues, etc. in future.

The new dependency is an updated version of an existing dependency (hibernate-types-52) - the developer renamed the library at some point. More info here: https://vladmihalcea.com/hibernate-types-hypersistence-utils/

Disadvantage: Limited array type support

Postgres array support is limited to primitive data types and String, that is int[], long[], float[], double[], boolean[], String[]

For example, given a FragmentRepository method such as:

void deleteByAnchorIn(Collection<AnchorEntity> anchorEntities);

Changing to an array of AnchorEntity will not work. The following will throw an exception:

void deleteByAnchorIn(AnchorEntity[] anchorEntities);

Rather, the method would have to be changed to use anchor IDs:

void deleteByAnchorIdIn(int[] anchorIds);

Side note: in the above case, if using auto-generated SQL from the method name, the SQL generated using the anchor ID will be simpler than when using AnchorEntity objects, as Hibernate will generate a JOIN between FragmentEntity and AnchorEntity in the first case, but will only use FragmentEntity in the last case. So while the types are more restrictive, it may result in more optimal SQL code.

Advantage: Code already using Postgres arrays can be simplified

The current FragmentRepository has code already The current code for has code using Postgres arrays to support "LIKE ANY" logic (using LIKE with a collection). However, it currently uses query parameters, and is thus limited to 32K parameters.

    @Override
    public void deleteByAnchorIdAndXpathLikeAny(final int anchorId, final Collection<String> xpathPatterns) {
        final String queryString = "DELETE FROM fragment f WHERE f.anchor_id = ? AND xpath LIKE ANY (array[:parameterPlaceholders])";
        executeUpdateWithAnchorIdAndCollection(queryString, anchorId, xpathPatterns);
    }

    // Accept security hotspot as placeholders in SQL query are created internally, not from user input.
    @SuppressWarnings("squid:S2077")
    private void executeUpdateWithAnchorIdAndCollection(final String sqlTemplate, final int anchorId, final Collection<String> collection) {
        if (!collection.isEmpty()) {
            final String parameterPlaceholders = String.join(",", Collections.nCopies(collection.size(), "?"));
            final String queryStringWithParameterPlaceholders =
                sqlTemplate.replaceFirst(":parameterPlaceholders\\b", parameterPlaceholders);

            final Query query = entityManager.createNativeQuery(queryStringWithParameterPlaceholders);
            query.setParameter(1, anchorId);
            int parameterIndex = 2;
            for (final String parameterValue : collection) {
                query.setParameter(parameterIndex++, parameterValue);
            }
            query.executeUpdate();
        }
    }

Using native Postgres arraysarray support, this code can be written directly in the FragmentRepository interface using Query annotation, without the 32K limitation:

    @Modifying
    @Query(value = "DELETE FROM fragment f WHERE anchor_id = :anchorId AND xpath LIKE ANY (:xpathPatterns)", nativeQuery = true)
    void deleteByAnchorIdAndXpathLikeAny(@Param("anchorId") int anchorId, @Param("xpathPatterns") String[] xpathPatterns);

    default void deleteByAnchorIdAndXpathLikeAny(int anchorId, Collection<String> xpathPatterns) {
        deleteByAnchorIdAndXpathLikeAny(anchorId, xpathPatterns.toArray(new String[0]));
    }

Comparison

• Both solutions have same performance (as determined using Groovy performance tests modified to take 100,000 items)More testing is needed to confirm if list partitioning drops performance with large datasets.
• Solution #1 is explicit using 32K batches, and so easy to understand what it does
• Solution #2 is has shorter code, but may be not as clear very obvious to the reader what it is used forSolution #2 is more flexible, adding array support unlocks more Postgres featuresdoes
  • We could add a comment linking to this ticket to explain the 32K limit and why arrays are used
• Solution #2 simplifies code in the cases when we need Postgres array support.
• It's possible to use both solutions, where most appropriate.