MongoDB | 2620209

Info

We discovered a strange phenomenon. After in-depth research, we found that it was a bug in the implementation of ScanningReplicaSetMonitor. First, prepare a single-shard cluster, with one primary and two secondary, and a script query secondary, like this ``` import pymongo from pymongo import MongoClient import time c = MongoClient("mongodb://xxxxx/admin?readPreference=secondaryPreferred") while True: for _ in c.db.coll。find(): pass ``` Then , let’s look at a series of common operations and the phenomena behind them. *  Set secondary 1 hidden = true ， a few seconds later，set secondary 1 hidden = false. at this time , I will find that only node 2 has query operation. And node 1 have a large replicaSetPingTimesMillis in mongos>db.adminCommand("getDiagnosticData").data.connPoolStats.replicaSetPingTimesMillis { "mongo109" : "x1:27017" : 2.459, "x2:27017" : 2.289, "x3:27017" : 9223372036854776 }, } Restart node 1 Then everything will return to normal, queries are distributed normally, and replicaSetPingTimesMillis is normal. Set secondary 1 and secondary 2 both hidden = true and a few seconds later revert to hidden = false ; at this time , I will find queries are distributed normally but all secondary replicaSetPingTimesMillis is large ; * After restart node 1 ,onle secondary 1 has query operation. The key reason behind the above phenomenon is that : * ServerPingMonitor::onTopologyDescriptionChangedEvent just remove monitors that are missing from the topology ； but don't add new monitors; int struct LatencyWindow , Due to the following code, there will be a Window (max(),max()) upper = (lowerBound == HelloRTT::max()) ? lowerBound : lowerBound + windowWidth; I think ServerPingMonitor is a bug, LatencyWindow is a feature

Top User Comments

Steps to Reproduce

a single-shard cluster, with one primary and two secondary, and a script query secondary Set secondary 1 hidden = true ， a few seconds later，set secondary 1 hidden = false. Restart node 1 Set secondary 1 and secondary 2 both hidden = true and a few seconds later revert to hidden = false ; Restart node 1