Experience of three meetings
-
The relationship between logstash and beans is that between the server and the client. Logstash takes up a lot of resources, while beans are lightweight; The official advice is that logstash should be deployed independently rather than mixed with business applications; -
The 6.5 version of Elastic search improves memory utilization and recycling even more, and supports java 11; The official said that after their users upgraded the version, the resources occupied by various parameters were much less.
-
China Mobile's ELK scale, business status, and methods to improve the utilization of machines, such as: deploying multiple nodes and multiple storage disks on a machine can improve the IO capability; -
The scenario is to collect logs to the ELK stack and connect 14 business lines. The log standard is their busy point at this stage, and standardization is important for any company; -
It also briefly talked about the use scenario of the K8S, encapsulating the configuration of logstash and business logs to the docker, and running on the K8S; -
Some simple optimization methods, briefly talking about the tool chain in the community; -
The different roles of the cluster are split. What's more amazing is that their client nodes are configured with 128G of memory.
-
Didn't listen. I went out to talk about technical details with the official staff; -
After a brief look at ECE, this sharing should be an official one-stop solution for operation and maintenance to promote enterprises.
Communicate with the official architect
-
At present, some indicators of our cluster are normal. For example, 22 nodes, 4257 partitions, 12.6BIL documents, 17.58T storage usage; -
The importance of timing and the constraints of search scope are also important; -
The maximum warm data processing capacity of a node is 1T data; Aggregation is a heavy operation. It will take out the index files recorded in all segments within the search range, read them from the storage disk to the memory, and then calculate; -
As for the cluster role, the master has very low requirements for resources, and some users will also put the master node on the K8S to play the advantages of hybrid cloud;
-
-
The division of focal group roles will improve the performance. He said that it may be 10%~20%, and some special business scenarios will even double the performance; -
In terms of fielddata, he said that generally only text search is used when keywords are highlighted, that is, we can reduce the proportion of fielddata, or even turn it off; -
I don't think he is very clear about the problem of _id. He gave a statement that it may be a problem of mapping, but _id is keyword by default, and there is no mapping; -
The real-time processing capacity is linear with the number of machines.
Improvement points of cluster
-
Cluster roles need to be split, and master nodes have low requirements for configuration; -
One machine can deploy multiple nodes, which can fully improve the utilization of the machine; -
The fielddata configuration can be minimized or even disabled; -
It is better for a machine with multiple disk slots. It would rather have less capacity of each disk than more disks. Usually, the business bottleneck is mainly on IO. Multiple disks can share the pressure of reading and writing files to IO.