Why did EcChronos connect to external datacenters, and does agent/master still require that? #1266
-
|
I’m testing the agent/master branch (not a released build) and noticed that the agent appears to no longer make outbound connections to external datacenters that older releases used to contact. Could you please confirm whether this is correct, and if so, explain why the external datacenter connections are no longer required for the agent, and why they were required previously? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
That was not a very good feature in old ecChronos, it did more harm then good. In the original implementation it was a concern that the locking mechanism would have consistency issues in larger clusters, it was believed that this could be avoided by routing locking requests to the datacenter that the lock was related to, so all locking requests related to a DC was processed in that DC. In real life it turned out that this concern was unfounded and no real consistency issues was noted. The big problem with this was that this implementation dose not scale well (or very badly), when the cluster expands the number of connections between ecchrons nodes and cassandra nodes grows alarmingly fast, the grows is probably worse then exponential and at some point it will start impacting cassandra nodes availability and if the cluster grows to big ecchronos can completely block the cassandra nodes. |
Beta Was this translation helpful? Give feedback.
-
|
Got it, thanks for the answer. |
Beta Was this translation helpful? Give feedback.
That was not a very good feature in old ecChronos, it did more harm then good.
In the original implementation it was a concern that the locking mechanism would have consistency issues in larger clusters, it was believed that this could be avoided by routing locking requests to the datacenter that the lock was related to, so all locking requests related to a DC was processed in that DC. In real life it turned out that this concern was unfounded and no real consistency issues was noted.
The big problem with this was that this implementation dose not scale well (or very badly), when the cluster expands the number of connections between ecchrons nodes and cassandra nodes grows alarmingly fast…