Insert Suspension During Helm Upgrade

[rere950303]

• milvus version: v2.4.20
• mode: distributed cluster

---

Hello,
I'm reaching out to ask about the safest way to upgrade Milvus using Helm without risking any data loss.

During the upgrade, I noticed that the datanode pods are rolled one by one.
In this case,

• should I manually call flush for each collection and stop all insert operations beforehand?
• Or is it safe to proceed with the upgrade without taking those steps?

[yhmo]

It is recommended that insert operations be stopped during the upgrade process. No need to call flush() since datanode automatically flushes data before restart.

[rere950303]

@yhmo
I see, thank you for your response.
If you don’t mind, could you please explain in a bit more detail why stopping insert is recommended?
Intuitively, it does seem safer to stop insert, but I’m asking because I’d like to understand more specifically what kind of issues might occur.

[xiaofan-luan]

All data is written to the WAL before it's even consumed by the DataNode.
However, you might have encountered some write errors during the upgrade process.
Using the K8s Operator is generally a better choice, as it provides fine-grained control over rolling upgrades.
In contrast, Helm lacks precise control over individual node behavior during the upgrade.

When you are doing rolling upgrade, milvus will also automatic flush to make sure recovery process is faster

[rere950303]

thanks all. really help for me 👍