envoyproxy/envoy

Avoid lossy LRS during cluster updates

Open

#12.823 aperta il 26 ago 2020

Vedi su GitHub
 (0 commenti) (0 reazioni) (0 assegnatari)C++ (5373 fork)batch import
area/load reportingbughelp wanted

Metriche repository

Star
 (27.997 star)
Metriche merge PR
 (Merge medio 8g) (378 PR mergiate in 30 g)

Descrizione

From https://github.com/grpc/proposal/pull/200, when a cluster update for X is received and the new cluster is active after warming, the old X cluster may still be draining for an indefinite period if it has active streams. Only the stats for new requests/streams on X will participate in LRS reporting at this point. This seems to be potentially misleading when reporting load only for the new X.

There are two scenarios to consider:

  1. When the EDS service name changes. I think in this case we should generate two LRS reports, one for old cluster X (including the older EDS service name) and another for new cluster X (including the new EDS service name).
  2. When the EDS service name is the same. I think a unified LRS report should be created aggregating old/new cluster X.

Guida contributor