eriq-augustine / elfs Goto Github PK

The current S3 readers/writers will perform requests over the wire one-for-one with requests made to the reader/writer.
However, we should probably make fewer, larger requests to S3.

This needs to be researched, and a good median found.

Right now, we use a more ACL-style group permissions.
We should probably just simplify and go with UNIX-style permissions everywhere.

This will lose some expressivity, but the ACL is really not necessary.

There are race conditions all over the place.
In pretty much every operation.

The real question is how we decide to handle them.
Should we:

• lock the driver on every command
• lock on the file level
• force the external caller to lock

At the driver level, the only critical concern is that that fat is not corrupted.
We will even tolerate things like incorrect access dates or counts as long as the filename and iv remains intact.

In most driver-level operations, we leak some information that the caller should not be privy to.
For example, when we make a dir we will first check to see if the parent directory exists and then we will check if we have permission to write there.
If the context user does not have read permissions on the parent directory, then they now have the additional information that the directory exists.

We do this so that we can return more clear errors.