Doing that at a startup would be easy, since you control the whole stack and can tune it to the use case, and almost certainly don't have problems of scale. It's 90% getting the technical stuff in place, assuming you have the budget.
Doing it at a place like Google is a different matter. The job isn't tuning a stack. It's coordinating multiple teams that own different services, and trying to make general-purpose services and frameworks work especially well for a specific case, even in the face of competing demands and standardization forces. It would be 90% political, and the task-specific differentiation required may simply be insurmountable - the extra wrinkles for your use case might cost more in complexity and reliability than it's worth.
What makes you think a Principal engineer is tasked with that? Principal engineers do mostly stuff like create new infrastructure technology like spanner, or machine learning frameworks, or similar stuff. Google doesn't have that many principal engineers, they correspond to director on the manager ladder, but there are way fewer principal engineers than directors as it is much harder to climb the contributor ladder than the manager ladder. Most Google products doesn't have a single principal engineer on the team.
Drive almost certainly has an L8 on the team somewhere. It is a big product. And it could be the case that "find a completely new indexing structure that makes drive search and folder navigation 10x faster" is a L8 task. Drive is indeed a lot slower than I'd personally like and a lot of that comes down to the way that indexing works when you cannot group users together, which is at the very core of the product.
Doing it at a place like Google is a different matter. The job isn't tuning a stack. It's coordinating multiple teams that own different services, and trying to make general-purpose services and frameworks work especially well for a specific case, even in the face of competing demands and standardization forces. It would be 90% political, and the task-specific differentiation required may simply be insurmountable - the extra wrinkles for your use case might cost more in complexity and reliability than it's worth.