What execution decisions forbid

What it forbids

You cannot revise a decision whose alternatives you have already closed, so adaptivity is bounded by what remains downstream.
You cannot observe without stopping, so every measurement costs a barrier.
You cannot bind a plan early and a parameter late without accepting the sniffing pathology. The four escapes manage it; none removes it.
You cannot cache a result without a cheap way to detect input change.
You cannot have both plan determinism and plan adaptivity as defaults.

The escapes and their prices

Escape	What it buys	Price
Barriers	Observability	Latency and space
Compilation	Per-tuple speed	A fixed cost that only a large n repays
Caching	The whole computation	Invalidation, and a bet on recurrence
Robust planning	The tail	The mean

Every one of these moves a cost to a different time, a different resource or a different failure mode. None makes it disappear.

Case, resolvedConstructed case · Ananya at Kiranamart

One store takes 9 minutes, the others 0.2 seconds

A parameterized store dashboard took 0.2 seconds for 390 stores and 9 minutes for the largest, and an ad hoc join planned for 5,000 rows received 40 million. Where it started: Execution decisions →

Problem 1 (the largest store takes 9 minutes). The cached plan was built when a small store's id was first used: an index lookup tuned for 200 rows. Reused for the largest store (millions of rows), it is a disaster, while the literal query, planned fresh, takes 6 seconds. Ananya has the four options. She chooses a per-class approach: the BI tool tags the largest 10 stores, and for those it forces a custom plan on every run; for the rest, it keeps the generic plan, which is robust. She also notices the timing: the first complaint came right after auto-analyze, which invalidated cached plans and flipped them. Now she checks plan changes at analyze time.

Problem 2 (the ad hoc join). The plan was made on an estimate of 5,000 rows. The first shuffle made the true size, 40 million, visible. With adaptive execution on, the engine turns the sort-merge join into a broadcast join if the other side is small, and splits the skewed partition. It helps only because the error was downstream of the first barrier. Had the error been an unprunable predicate in the first scan, adaptivity could not have helped.

She also puts a result cache on the "Store sales this week" tiles: the table has a snapshot id, so each hit costs a version check, and a refresh of the nightly load invalidates cleanly.