share layout: think about round trips

[Imported from Trac: page NewImmutableEncodingDesign, version 5]
warner 2010-01-07 01:48:32 +00:00
parent 61984d5585
commit f39ab52979

@ -37,6 +37,17 @@ This page is for designing the encoding format for these new immutable files.
(servers can compare share contents against the UEB, and we could put a (servers can compare share contents against the UEB, and we could put a
copy of the UEB hash into the share, but servers would continue to be copy of the UEB hash into the share, but servers would continue to be
unable to make sure the share was in the right place) unable to make sure the share was in the right place)
* make the share layout more deterministic, to get rid of the RTT that's
needed to fetch the offset table before we can fetch the UEB. Maybe
reserve space for the UEB at the start of the share (but that makes it hard
to have an extensible open-ended UEB).
* since we probably do need to stick with the offset table, see if there's a
way to guarantee having the same offsets across all shares. This would allow
the table pulled (and validated) from one share to be used to compute read
vectors for all other shares. We probably already have identical tables,
but it'd be nice to make the guarantee explicit. The tables need to be
covered by the UEB, else we can't share them: the first read might be from
a corrupted share, and that would malign (hah!) the other shares.
# Options # Options