We have up to 3 layers of retries for telemetry-based autotests.

We catch telemetry exceptions and retry if login fails:
https://cs.corp.google.com/chromeos_public/src/third_party/autotest/files/client/common_lib/cros/chrome.py?l=170

JOB_RETRIES in most autotests:
https://cs.corp.google.com/android/external/autotest/client/site_tests/login_LoginSuccess/control?l=12

And we retry the smoke suite in VMs if a single test fails.

These policies exist in trybots, PFQ bots, CQ bots and the main waterfall.

Increases in flakiness rates end up being ignored, and we tolerate a high degree of variability in how long a bot runs take.