Try fixing e2e network flakiness

[jtoar]

This is an attempt to fix network flakiness using the strategy in vercel/next.js#28264. I tried this out on my fork of redwood and it seemed to actually work (PR here: https://github.com/jtoar/redwood/pull/155), but I have no idea what this really does. (The original comment says that it "disables TCP/UCP offloading", but I don't really know what that entails.)

[thedavidprice]

@jtoar This is definitely interesting, especially given that 1) we've had external networking issues in the past (npm package installation) and 2) we don't have any visibility/control over how the services/containers are networked within the workflow, which could definitely be causing issues given that we spin up processes outside of other services.

Let's give it a shot!

Research link:

• https://linux.die.net/man/8/ethtool

tx off rx off
Related to network receive/transfer and turning off the default pause. Seems reasonable to try.

Looks like we'll need to run some tests regarding eth0. See:

• https://linuxhint.com/ethtool_commands_examples/
• Section "Displaying Network Interface Card Information with ethtool"

Not sure if we can assume unique name is same each time without further research. Will need to run sudo ip link show within a workflow to verify. And might need to use the return output to pass sudo ethtool -K <NIC_name> tx off rx off

[thedavidprice]

If this works, we should add to add workflows, setting a condition to run in case of Linux only.

[jtoar]

Ok! I'll merge and rebase some of the existing PRs that are failing to see if this needs any more of the tweaks you mentioned.

[jtoar]

@thedavidprice Here's the output of sudo ip link show:

Maybe we need to tune docker0 too?

       # For network flakiness. See https://github.com/vercel/next.js/pull/28264
       - name: Tune linux networks
         run: |
           sudo ethtool -K eth0 tx off rx off
+          sudo ethtool -K docker0 tx off rx off