Skip to main content
Solved

Cribl Stream: Firewall log aggregation issues with multiple workers

  • Nov 4, 2026 11:52 AM UTC
  • 1 reply
  • 0 views

msallam
This message originated from Cribl Community Slack.
Click here to view the original link.

Hello, Looking for a sanity check on a firewall log aggregation issue I ran into. Setup: Cribl Stream, distributed on-prem, 2 primary FW workers running an Aggregations function to correlate Fortigate FW logs before sending to Splunk. Spec I was aiming for: aggregate events sharing Source IP, Destination IP, Source Port, Destination Port, and Action within a 1-2 second window, with an Event_Count field showing how many raw events were combined. While validating against the original raw index, I found a case where my aggregated output showed Event_Count = 2 for a given 5-tuple, but the raw index actually had 11 matching events — with a burst of at least 3 events all within a few seconds of each other (well under any reasonable gap threshold), not 10+ seconds apart as I'd assumed. Root cause I landed on: two compounding issues — 1. My Aggregations Time window was set to 1s, way too short relative to a multi-second continuous burst, so a single burst was getting chopped into several tumbling windows. 2. The 2 primary workers each run Aggregations independently (shared-nothing per Worker Process), splitting one correlated burst into partial aggregates across both workers, with no merge step downstream. My fix: consolidate FW log routing so all events for these flows hit a single worker (removing the cross-worker split), and set Time window: 2s, Lag tolerance: 5s, Idle bucket time limit: 10s, Flush on stream close: OFF. Question for the community: for firewall log correlation like this — same 5-tuple, tight time proximity — is single-worker consolidation + tuned tumbling window the standard approach, or does anyone handle this differently? Want to make sure I'm not building around a limitation there's a cleaner native pattern for. Thanks!

Best answer by Oliver Hoppe

Hi Marwan, even if you route the events just to a single worker that single worker has multiple worker processes which aggregate the events independently. Stream as of today cannot aggregate across multiple worker processes, as you correctly stated is using a shared nothing software design. Especially for low volume data types the savings might not be really satisfying. For huge datasets with a bigger aggregation window of 30-60 seconds and not too high cardinality (5-touple might have already a pretty high cardinality) will show most likely a good reduction assuming you are not using hundreds of worker processes.
View original

1 reply

Oliver Hoppe
  • Employee
  • Answer
  • Nov 4, 2026 11:52 AM UTC
Hi Marwan, even if you route the events just to a single worker that single worker has multiple worker processes which aggregate the events independently. Stream as of today cannot aggregate across multiple worker processes, as you correctly stated is using a shared nothing software design. Especially for low volume data types the savings might not be really satisfying. For huge datasets with a bigger aggregation window of 30-60 seconds and not too high cardinality (5-touple might have already a pretty high cardinality) will show most likely a good reduction assuming you are not using hundreds of worker processes.