Introduction
On graphd/storaged reboots/failures, we observed that QPS skews toward remaining living servers, and QPS are never balanced again without intervention(client reboot, network fluctuation, etc).
The more servers we have, the lower the probability that a new session is created on a recently started server, because pos is only an index.
This undermines the purpose of a distributed database.
Proposal
Implement a new priority queue based on existing connections count for graphd server to replace pos, enabling each client to balance connections across all available servers. (I believe this would serve as a suitable default configuration.)
Introduction
On graphd/storaged reboots/failures, we observed that QPS skews toward remaining living servers, and QPS are never balanced again without intervention(client reboot, network fluctuation, etc).
The more servers we have, the lower the probability that a new session is created on a recently started server, because pos is only an index.
This undermines the purpose of a distributed database.
Proposal
Implement a new priority queue based on existing connections count for graphd server to replace
pos, enabling each client to balance connections across all available servers. (I believe this would serve as a suitable default configuration.)