The encoding field says it is a hash, and the type field records whether the layout is compact or a full table.

The object type field says the value is a hash, and the encoding field says whether it is stored compactly or as a dictionary.

A single representation field covers both, because hash-compact and hash-table are two distinct object types.

The key itself carries a type prefix, and the value structure only needs to record its encoding, which keeps the value structures small and uniform.

It is invisible because the upgrade happens during a fork, where clients cannot observe process memory.

It is invisible only if a client never rereads the key, since older readers still see the compact form until they reconnect.

It is invisible because Redis rewrites client replies to imitate the compact encoding's field ordering.

Encoding changes are internal because the logical type and command behavior stay the same while the physical representation changes.

LPUSH against a key holding a string returns a wrong-type error because the key exists but its logical object type is not a list.

The string is converted into a single-element list and Ada is pushed in front of the original value, mirroring the way APPEND grows a string value in place.

LPUSH overwrites the key with a fresh list, because write commands take precedence over existing values.

The push succeeds at the encoding level, and the key reports type list from the next access onward.

The reference count drives eviction since heavily shared objects survive, while the encoding field drives validation.

The TTL field drives eviction for every policy, while LRU metadata validates that commands access keys in a safe order.

LRU or LFU metadata helps eviction policy, while the logical type field helps command validation.

Eviction and validation both read the encoding field, because compact values are cheaper to evict and faster to check.

Raw bytes would be smaller and faster, so logical type exists in the file format mainly for debugging tools.

Logical type only matters because raw pointers would dangle after restart, and a relocating loader could fix those without any type information.

Persistence needs logical type only when AOF and RDB are combined; a pure RDB file stores raw memory pages.

Persistence needs logical type so recovery can restore behavior and encodings, not just copy opaque process memory bytes.