Part 2 of 5 · 1 min read
The problem: an ever-growing log means ever-growing recovery
Without a stopping point, recovery would need to replay the log all the way back to the very beginning.
Think of trying to reconstruct today's events by reading a diary from its very first page every single time, even after years of entries have already piled up.
In plain words: without some marker of "everything before this point is already safely reflected on disk," crash recovery would have no way to know how far back it actually needs to look. In the worst case, it would need to redo and undo starting from the very first log entry ever written, which would make recovery time grow without bound as a database accumulates more and more history. A checkpoint exists specifically to give recovery a safe, recent starting point instead.
A worked example. A database that has been running for years, with millions of log entries, would face an extremely long recovery process after any crash if it had to replay all the way back to its very first log entry every time.
The key lines. No new SQL; this growing-recovery-time problem is exactly what checkpointing, covered next, is designed to solve.
Unbounded recovery problem
The unbounded recovery problem is the situation where, without a recent safe starting point, crash recovery would need to replay the log all the way back to its very beginning, making recovery time grow without bound over a database's lifetime.
Example: reconstructing today's events by rereading a diary from its very first page every time, instead of starting from a recent, known-good point, mirrors the unbounded recovery problem.