Hi <!here> I have a 3

storaged

host based nebula graph cluster deployment, which has a large amount of data. As per the documentation: https://docs.nebula-graph.io/3.4.1/5.configurations-and-logs/1.configurations/4.storage-config/#storage_configurations_for_large_dataset I have set

enable_partitioned_index_filter=true

in every storage nodes

nebula-storaged.conf

. Although 2 hosts use low amount RAM (expected behavior), the third host overloads the RAM and crashes. Please suggest if there is anything missing to get the cluster working with low RAM allocations.

Checking leader distribuition first to make sure it's balanced may help.And check rocksdb's block cache. Please let me know your data size， and the crashed storage's memory usage.

The space has 100 partitions.

+--------------+------+-----------+----------+--------------+---------------------+------------------------+---------+
| Host         | Port | HTTP port | Status   | Leader count | Leader distribution | Partition distribution | Version |
+--------------+------+-----------+----------+--------------+---------------------+------------------------+---------+
| "10.0.1.8"   | 9779 | 19779     | "ONLINE" | 33           | "MDB:33"            | "MDB:33"               | "3.3.0" |
| "10.0.1.95"  | 9779 | 19779     | "ONLINE" | 33           | "MDB:33"            | "MDB:33"               | "3.3.0" |
| "10.0.1.252" | 9779 | 19779     | "ONLINE" | 34           | "MDB:34"            | "MDB:34"               | "3.3.0" |
+--------------+------+-----------+----------+--------------+---------------------+------------------------+---------+

Vertex and Edge Counts:

| "Space" | "vertices"    | 4360152   |
| "Space" | "edges"       | 195535793 |

Disk usage is around 50G each across hosts. RAM usage shoots over 250Gb.

That's weired. What kind of queries were you excuting then? A skewed big query is possible but would hardly happen since data is hashed now. You can also check current config from http interface to see if it's expected.

I am not yet executing any queries. My concern was with respect to the start up. I was expecting that

enable_partitioned_index_filter=true

will reduce the RAM requirement, which performing as expected in 2 hosts of the cluster (Max. RAM usage: 10-12gb), but the host with

Leader Count 34

tries to use more than 250Gb.