Vector DB node OOM crashes during heavy embedding indexing

Question

Vector DB node OOM crashes during heavy embedding indexing

Jane Smith 0

We are running into a critical issue in our corporate environment where our data ingestion pipeline is crashing the vector database nodes whenever we try to index high-dimensional embeddings. The host system's memory gets completely saturated, which ends up taking the node offline and killing the ingestion job.

How can we properly tune the segment size and configure the disk offload parameters to stabilize indexing and prevent these OOM crashes? Any guidance on best practices for high-throughput pipelines would be greatly appreciated.

0 comments

1 answer

Your answer

Answer 1

Hello Jane Smith,

The out of memory crash occurs because building graph based indices such as HNSW for high dimensional vectors requires several times more RAM than the raw data footprint. When segment sizes are left too large while multiple ingestion threads run indexing jobs simultaneously, host memory reaches full saturation, which forces the operating system to terminate the vector database process.

To resolve this, reduce the maximum segment size in your database configuration down to 256MB or 512MB instead of the default 1024MB, which effectively caps the peak memory required during background consolidation and index merges. At the same time, enable memory-mapped files via mmap for your vector fields and configure your disk offload thresholds to aggressively flush in-memory buffers when system RAM utilization hits 70 to 75 percent. Additionally, limit the number of concurrent indexing worker threads to align strictly with your physical CPU core budget to prevent memory runaway. If this configuration stabilizes your data pipeline and resolves the OOM crashes, please accept the answer.

Tracy Le.