Elasticsearch version 8.19.1

Auto-sharding for data streams caused unsupported replica scaling when the index mode was set to LOOKUP. This happened because lookup mappers do not support scaling beyond one replica. #131429 resolves this issue by disabling auto-sharding for data streams with LOOKUP index mode, avoiding unsupported replica settings.

Parsing EQL queries failed when comparing the same IP-mapped field to multiple values joined by an OR expression. This occurred because lookup operators were internally rewritten into IN expressions, which are unsupported for IP-type fields. #132167 resolves the issue and ensures EQL can now successfully parse and execute such or queries involving IP fields. (issue: #118621)

Inconsistent equality checks caused constantNullBlock.equals(anyDoubleBlock) to return false, even when doubleBlock.equals(constantNullBlock) returned true. This asymmetry led to unreliable comparisons and mismatched hashcodes when ConstantNullBlock was functionally equivalent to other standard blocks. #131817 resolves the issue and ensures both equality and hashcode functions are symmetric for these block types.

A ConcurrentModificationException caused test failures in CrossClusterAsyncEnrichStopIT.testEnrichAfterStop under certain conditions. This happened because the ES|QL driver added a live operator list to the DriverStatus object, which could be modified while the status was being serialized. #132260 fixes the issue by copying the operator list before storing it, preventing concurrent changes during status reads. (issue: #131564)

Calling the to_lower or to_upper functions with no parameters caused a null pointer exception (NPE), instead of returning a clear error. This behavior was a result of an older implementation of these functions. #131917 resolves the issue and ensures that empty parameter calls now return the correct error message. (issue: #131913)

Sorting across multiple indices failed when one index contained an aggregate_metric_double field and another did not. In this case, the missing field was encoded as NullBlock but later incorrectly decoded as AggregateMetricDoubleBlock, which expects four values. This mismatch caused decoding errors. #131658 resolves the issue and also improves mv_expand by returning the input block unchanged for unsupported AggregateMetricDoubleBlock values, avoiding unnecessary errors.

Combining two BulkResponse objects with ingest_took set to NO_INGEST_TOOK resulted in a combined ingest_took value of -2, which was invalid. This occurred because the combination logic failed to preserve the sentinel NO_INGEST_TOOK constant. #132088 resolves the issue and ensures the result is correctly set to NO_INGEST_TOOK when applicable.

Loading large rows from a single segment occasionally created oversized pages when decoding values row-by-row, particularly for text and geo fields. This could cause memory pressure or degraded performance. #131053 resolves the issue by estimating the size of each page as rows are loaded. If the estimated size exceeds a configurable jumbo threshold (defaulting to one megabyte), row loading stops early, the page is returned, and remaining rows are processed in subsequent iterations. This prevents loading incomplete or oversized pages during data aggregation.

Grants the server module read/write access to the deprecated path.shared_data setting. #131680 resolves issues surfaced in internal testing and ensures compatibility with legacy configurations.

When using the simulate ingest API with a mapping_addition, the system incorrectly ignored the existing mapping of the target index and instead applied the mapping from a matching index template, if one existed. This caused mismatches between the index and simulation behavior. #132101 resolves the issue and ensures that the index’s actual mapping is used when available, preserving consistency between simulation and execution.

A recent refactor introduced a bug that caused the trained model memory estimation to double-count the number of allocations, leading to inflated memory usage projections. #131990 resolves the issue by reverting the change and restoring accurate memory estimation for trained model deployments.

A decoding error occurred when field names in _ignored_source contained non-ASCII characters. This happened because String.length() was used to calculate the byte length of the field name, which only works correctly for ASCII characters. #132018 resolves the issue by using the actual byte array length of the encoded field name, ensuring proper decoding regardless of character encoding.

Indices created prior to version 7.14 that used an index sort on a date_nanos field could not be opened in more recent versions due to a mismatch in the default index.sort.missing value. A change in version 7.14 modified the default from Long.MIN_VALUE to 0L, which caused newer versions to reject those older indices. #132162 resolves the issue by restoring the expected default value for indices created before 7.14, allowing them to open successfully in newer versions. (issue: #132040)

The timeout cancellation callback registered in QueryPhase via addQueryCancellation was not removed after the query phase completed. This caused unintended timeouts or cancellations during subsequent phases under specific conditions (such as large datasets, low timeouts, and partial search results enabled). #130279 resolves the issue and ensures predictable behavior by reintroducing the cleanup logic. (issue: #130071)