Or Our mapping and enrichment activities read large numbers of individual RDF aggregations from Marmotta, modify them, and save them. This results in high volumes of UPDATE and INSERT queries being made to PostgreSQL. Especially when we run multiple mappings or enrichments at the same time, we see significant amounts of lock contention. With shared buffers being dirtied at a high rate, we have also seen frequent activity to flush the buffer pool to disk, which is accompanied by some degree of I/O wait. We've tried to arrive at a balance of frequent-enough vacuuming and analysis, without causing too much I/O wait while this happens. We are still experimenting with the best settings for locking, checkpoints, and autovacuuming.
...