Memory management in Yandex Managed Service for Valkey™
For user data on the Yandex Managed Service for Valkey™ cluster hosts, 75% of the total amount of RAM is allocated. The remaining RAM is committed to the Valkey™ process. You can reduce the amount of RAM allocated for user data by using the Maxmemory percent parameter. This may be helpful if an out-of-memory error occurs in the cluster.
By committing RAM on the host, you can:
Improving the performance and stability of the Valkey™ process
Yandex Managed Service for Valkey™ cluster hosts use the vm.overcommit_memory = 0 Linux kernel setting. The Valkey™ process leverages it to minimize the number of swap file accesses. This increases performance and ensures stable operation of the Valkey™ process, for example, during replication and backups.
Backup stability
Backups in Yandex Managed Service for Valkey™ are created based on a consistent snapshot of the process memory image. The snapshot is created by copying the original Valkey™ process through the fork() system call.
With fork(), you can reduce RAM usage, as the original Valkey™ process and its copy share the same memory pages. The committed RAM and the vm.overcommit_memory = 0 setting make fork() more secure, since creating copies does not affect the main Valkey™ process performance.
When fork() is running, it uses the Copy-on-Write feature: when a shared memory page is modified, its copy is created to write the changes to. This is why, if you write anything to the cluster during backups, the updated and new data take up more space in terms of memory. This means, when data writing is intensive, you might run out of RAM on your hosts even if you have the committed RAM. In this case, the host operating system will terminate the Valkey™ processes with the out-of-memory (OOM) error. Valkey™ will become unavailable on this host, and the replication and backup operations will be interrupted. To avoid this:
- Make sure to start backups when the cluster load is minimum.
- Increase the amount of RAM by upgrading the host class.