A cloud file still needs a home

When a file appears in a folder, that does not tell you where its bytes live.
A local file may occupy the internal disk. A placeholder may contain a name,
a size and a reference to an object stored somewhere else. The application
must retrieve the bytes before it can use them. A fast network can hide some
of this wait, but a disconnected computer cannot fetch a missing block.

Video players can read a small part of a recording and start showing frames.
They need enough container metadata to find the requested section. An editor
that scans an entire recording can cause a very different access pattern.
Scrubbing backwards and forwards also changes which blocks are useful.
Measure cold opens, repeated opens and seeking separately.

Compression replaces predictable data with a shorter description. Repeated
words, recurring field names and similar nearby pixels provide useful clues.
The decoder needs the same dictionary, prediction rules or model weights.
Saving those rules somewhere else does not make their storage cost disappear.
An unfamiliar input may leave little for the model to predict.

Lossless archives restore every original byte. A different image encoding can
restore the same pixels without restoring the source file's metadata or byte
layout. A lossy encoder changes the signal to reduce its size further. These
three guarantees answer different questions. State which one a test checks.

An encrypted object should hide the patterns in the original file. Compress
the plaintext first when the design permits it, then encrypt the result.
Keep the key separate from the stored ciphertext. Recovery procedures matter
because a lost key can make an otherwise intact archive unusable.

A storage quota is a promise about capacity, rather than a measurement of
bytes used today. Compare monthly subscriptions using the same billing term
and currency. Keep the cost of restoring an archive in the calculation.
Ask how long an export takes on the connection that will actually carry it.

For this test, the samples are deliberately small and synthetic. A smooth
gradient favours prediction. Random noise frustrates it. A production archive
contains a mixture whose balance depends on its owner. Test your own mixture
before turning this example into a purchasing or infrastructure decision.
