Memgraph v3.2.0 - Apr 23rd, 2025

⚠️ Breaking changes

• Procedure results are now populated with null values by default. Procedures no longer throw if all result fields don't get inserted; instead, they return null in those fields. #2790

✨ New features

• Added support for composite indices. Composite indices are indices on labels and two or more properties. Composite indices significantly improve database performance on queries that filter against multiple properties. To take advantage of this improvement, users should create composite indices for frequently queried combinations of properties. #2760, #2887
• Added metrics tracking related to replication and high availability. This includes metrics like the number of RPC messages sent, the duration of the recovery etc. Users can use our prometheus-exporter to get immediately cluster info status. #2772
• Added mgp_result_reserve function to the C-api for reserving memory for procedure results. #2790
• Added edge TTL. The system can now automatically delete stale edges using the specified property ttl inside the edges. #2730
• Added global index on edge properties. Since edges have only one relationship type, the global property index will have a data structure that is tracked over properties with the syntax CREATE GLOBAL EDGE INDEX ON :(prop). #2730
• Expanded storage access types to UNIQUE, READ, WRITE and READ_ONLY. A better differentiation of query types allows more queries to run in parallel. Users can now run read queries while creating snapshots, even in ANALYTICAL mode. 2798
• Added support for OR label expressions, enabling queries like MATCH (n:Label1|Label2) to retrieve nodes with any of the specified labels. If indexes exist on the labels, the planner rewrites the query as a UNION of index scans for efficient execution. Additionally, simple WHERE clauses with OR label checks are optimized similarly, reducing reliance on full scans. #2783

🛠️ Improvements

• There is now an improved logic for selecting the new MAIN instance during the failover in the MT environment. An instance that has the newest databases is considered the newest. Users will now have a smaller chance of experiencing the loss of committed data #2812
• Metrics related to HA coordinators are now immediately reset after the client pulls them. Users can now see the aggregated value from all coordinators when using Memgraph's Prometheus exporter. Also, the data won't be lost after the coordinator's restart. #2802
• The memory footprint of TypedValue (internal data structure used mostly during query execution) got smaller. The optimization results in lower memory usage in query modules that return many results (e.g., modules that return the whole graph, like PageRank) by a few percent. #2806
• Reduced memory footprint of mgp_result_record used for returning results in all procedures APIs. The optimization reduces memory usage by up to 50% in query modules, which return many results, such as the whole graph (e.g.,PageRank). #2790
• utils::Scheduler now skips backed-up tasks. Background tasks that took longer than the period will not be called multiple times after the execution. It lowers the overall workload and frees up resources. 2853
• Schema info edge updates made faster. Previously, edge deletions would cause a slow commit. Users can now use the run-time schema info more freely without impacting the performance. 2844
• Snapshot creation has been parallelized. Edges and vertices are batched (as defined via flag). These batches are processed in parallel on a number of threads. Users can now create snapshots in less time, freeing up resources for other tasks. 2818

🐞 Bug fixes

• Parsing a query in which a pattern comprehension is nested (such as RETURN [[(x)-->(y) | x]];) no longer causes a crash. #2716
• Replica won't crash anymore when performing recovery and when disk access to the durability folder fails. Instead, it will return an empty response, signaling the main that the recovery isn't completed. Users should expect the same behavior as before. #2780
• The scheduler (utility module) will release its lock before starting to execute its function. This will enable deadlock-free concurrent execution between the periodic snapshot thread and the thread that performs demotion under replication. #2882
• During the promotion of the MAIN instance, the promotion will be aborted if registering one of the replicas fails. In that case, the coordinator will retry the promotion after a few seconds. There is now a smaller chance of users experiencing failed failover. Previously, the RPC communication would fail, and MAIN would partially register some of the replicas. #2869
• The newly elected coordinator leader will perform a failover correctly when the leader inherits the cluster state without the main instance. #2857
• When manually promoting an instance, it is now unnecessary that all replicas are alive. Users can now use the SET INSTANCE instance TO MAIN query in more situations. #2857
• If a snapshot gets created while MAIN receives a request to demote itself to REPLICA, the snapshot creation will get aborted. Additionally, the thread creating snapshots will be paused on demotion. Similarly, when the REPLICA is promoted to the main, the snapshot thread will be resumed. #2846
• The DUMP DATABASE query generated the wrong query for ANY triggers. Dumping the database now generates the correct trigger query. 2855
• Fixed missing edge property index and unique constraint garbage collection. No slow memory creep up while using constraints or edge index. 2850
• Metadata from WAL files was unnecessarily used during recovery, even for old WAL files that shouldn't have been used. This could've caused a possible data loss because newer WAL files wouldn't be used in some situations while recovering. #2892
• DUMP DATABASE has more stability guarantees. The output order has changed for better memory usage and performance, and a sort is applied to the type constraints. The durability write bug in v2.19.0, v2.20.0, v2.20.1 for the 3D Cartesian data type was fixed in v2.21.0, but in a way that reading durability from those versions would no longer be possible if a "corrupt" 3D Cartesian was written into those durability files. Those durability files can now be read, and the engine handles the "corrupt" values. #2888
• Fixed unreliable authentication deserialization. Users can now reliably upgrade auth data between different versions, as well as between the community and enterprise. 2915