Comparison
In-place upgradevsSnowflake server
In-place upgrade
you ran the package upgrade on the running fleet, and now the machines differ from each other by whatever failed halfway.
Modifying running machines rather than replacing them. It is faster and it preserves anything local, which is exactly why it accumulates divergence: every partial failure leaves a machine in a state no image describes. It remains the only option for genuinely long-lived stateful hosts, which is where the discipline of configuration management still earns its keep.
Full entry →Snowflake server
there is one machine everyone is afraid to reboot, and its configuration exists only in the history of who logged in.
A host whose configuration has diverged so far through manual changes that it cannot be reproduced. It is the direct output of in-place changes over years, and the risk is not that it is unusual but that it is unrecoverable — a hardware failure becomes an archaeology project. The standard remedy is to rebuild it from code once, deliberately, while it is still working rather than after it is not.
Full entry →