Allkeys policies can remove a non-TTL key, while volatile policies cannot because they only choose among keys with expiration metadata.

Only volatile policies could, because they fall back to non-TTL keys whenever no TTL keys remain to evict, logging a warning each time the fallback widens the candidate pool.

Any policy could have done it, since over maxmemory Redis ignores policy scope and frees whichever keys are largest.

Only noeviction could, because it removes keys without TTLs specifically to protect data that is set to expire.

Sampled LRU tracks hot keys better because sampling naturally favors the keys it sees most often.

Redis uses exact LRU for small keyspaces and switches to sampling automatically above a size threshold.

Exact LRU observes every access, but sampling is cheaper and usually good enough for Redis latency and memory overhead goals.

Exact LRU would require pausing writes during its list maintenance, which is why no production system implements it.

Redis evicts the least-recently-used key anyway, because staying available outranks the policy's scope.

The write succeeds and maxmemory rises automatically by a configured overshoot factor.

The write blocks until some key receives a TTL or memory is freed by another code path, waking periodically to retest whether the allocation now fits.

The write should receive an OOM error because volatile-lru has no legal eviction candidate when no keys have TTLs.

All memory looks identical to the allocator, so enforcement needs only the single global used-bytes counter.

Value memory, client output buffers, and replication backlog are owned by different subsystems, so enforcement must account for more than key payload size.

Output buffers and the backlog are allocated outside the Redis process, which puts them beyond what maxmemory can observe.

The complication is double-counting, since values referenced by output buffers would otherwise be freed twice.

All commands including reads should be denied, since even GET allocates a reply buffer to serve its result.

Only commands on keys without TTLs should be denied, steering new writes toward data that can be evicted.

Commands that allocate or grow memory should be denied when eviction cannot free enough memory; read-only commands can usually continue.

DEL and expiration activity should be denied as well, freezing the keyspace until an operator intervenes — better a frozen dataset than one mutating under pressure.