Concerns that shape my own decision to run Marginalia Search as a service based architecture rather than one blob include:

* NUMA optimization - on multi-CPU architectures, I/O and/or memory bottlenecked software takes a pretty significant performance hit if you let is run on both CPUs, and this has knock on effects that also degrade the performance of other software running on the same machine.

* Network namespaces - if you're e.g. running an internet crawler, and want to assign one host machine multiple public IP addresses, this is much easier if you use something like ipvlan.

* Stateful long running process lifecycles (e.g. crawling) that coexist with the need for live redeployments of other parts of the system.