Building a Minimal Cluster

This topic has been translated from a Chinese forum by GPT and might contain errors.

Original topic: 最小集群搭建

| username: TiDBer_PHgBQFC6

【TiDB Usage Environment】Production Environment
【TiDB Version】
【Encountered Issues: Problem Phenomenon and Impact】
【Resource Configuration】Cluster built with 5 machines, each with 64GB RAM and 32 cores
May I ask everyone, how to plan the cluster more safely and reasonably?
Without needing Tiflash, how to plan it better?
Can TiDB and PD be configured on only 2 machines?

| username: 像风一样的男子 | Original post link

It is best to have three PDs, two TiDBs are fine, and three TiKVs.

| username: TiDBer_PHgBQFC6 | Original post link

Boss, there are only 5 machines:
tidb_server: 2 machines 101, 102
pd: 3 machines 101, 102, 103
tikv: 3 machines 103, 104, 105
monitoring: 1 machine 105
Is this suitable?

| username: 像风一样的男子 | Original post link

You can install three PDs and three TiKVs together, and install two TiDBs and monitoring together. The two TiDBs need to be load balanced.

| username: zhanggame1 | Original post link

If the business volume is large, it is recommended to have more TiDB nodes, as it is quite performance-intensive. Increase it to 4 nodes. PD does not consume much resources, so install 3 or 5 nodes as the documentation requires an odd number. For TiKV, 4 or 5 nodes are recommended, but I think 4 nodes are sufficient. Use the remaining machine for TiUP and monitoring.

| username: TiDBer_PHgBQFC6 | Original post link

There are only 5 machines in total. According to your implementation:
tidb 101, 102, 103, 104
pd 101, 102, 103
tikv 101, 102, 103, 104
monitoring 105
Will this ensure no issues?

| username: redgame | Original post link

It looks feasible, let’s give it a try.

| username: 有猫万事足 | Original post link

TiDB and PD can be configured on just 2 machines.

However, your machines are quite powerful, so you shouldn’t have to compromise. By using NUMA to bind and isolate CPU resources, you can achieve hybrid deployment.

Only machines with a single NUMA core need to consider deploying PD and TiDB on one machine and TiKV on a separate one. Without NUMA isolation of CPU resources, TiKV can max out the CPU under heavy write pressure, causing PD on the same machine to become unresponsive, which can crash the entire cluster.

As long as CPU resources can be isolated, there is no such concern, and they can be deployed together.

There are templates in the documentation, but you’ll need to calculate and adjust the parameters yourself. Given your machine’s performance, prioritizing hybrid deployment and switching only if it doesn’t work seems more appropriate.

| username: Kongdom | Original post link

Classmate, be ruthless: 1 TiDB, 1 PD, and 3 TiKV :thinking:

Or 1 server with 3 TiDB, 1 server with 3 PD, and 3 TiKV. Actually, I don’t think it’s necessary to do it this way.

| username: zhanggame1 | Original post link

No problem, and adjusting components in a TiDB cluster is very simple. You can scale up or down with just one or two commands.

| username: tidb狂热爱好者 | Original post link

tidb_server: 2 units (101, 102)
pd: 3 units (101, 102, 103)
tikv: 3 units (103, 104, 105)
monitoring: 1 unit (102)

| username: Kongdom | Original post link

PD recommends having an odd number of nodes.

| username: 有猫万事足 | Original post link

I only have 2 PDs purely due to lack of resources, running a small operation. :joy:
I don’t feel secure with just 1, and I can’t fit 3. Even so, the monitoring is still sharing a machine with other services.

| username: Kongdom | Original post link

:joy: I’ve been burned by even-numbered node split-brain issues~~~

| username: system | Original post link

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.