~bzr-pqm/bzr/bzr.dev

1185.1.29 by Robert Collins
merge merge tweaks from aaron, which includes latest .dev
1
Use of hashes in Bazaar-NG
2
**************************
3
4
* http://infohost.nmt.edu/~val/review/hash.html
5
* http://infohost.nmt.edu/~val/review/hash2.html
6
7
The main attraction of hashes in bazaar-ng is as an easy way to get
8
universally-unique IDs, or at least with a low chance of collision:
9
10
The first paper is a bit paranoid; the second has some sensible
11
advice:
12
13
1. Will compare-by-hash provide significant benefit -- save time,
14
   bandwidth, etc?
15
16
2. Is the system usable if hash collisions can be generated at will?
17
18
3. Can the hashes be regenerated with a different algorithm at any
19
   time?
20
21
We should try to abide by these rules.  I think they are possibly too
22
paranoid -- a real break of SHA-1 would have much wider security
23
implications -- but if a design that respects them is practical, it
24
should be preferred.
25
26
The first is probably true; the third is just a matter of making sure
27
we allow for the choice of hash to be varied in the format.
28
29
There are actually two variations on the second:
30
31
2a. Is the system safe if an attacker can generate hash collisions?
32
33
2b. Is the system safe if a user's own files contain collisions.
34
35
Regardless of cryptographic weakness, SHA-1 is unlikely to
36
"accidentally" collide, but it's possible that someone will
37
intentionally generate collisions (in research on SHA) and then want
38
to store them.  It would be unfortunate if that did not work.
39
40
An advantage of naming by hash is that it lets us store only a single
41
copy of identical files, but we have already decided__ that disk space
42
is pretty cheap.  It is perhaps enough to have a single copy of files
43
that do not change from one tree revision to the next.
44
45
__ costs.html
46
47
As far as an attacker: we will not automatically trust that ids from
48
one branch have the same value in another.  It is possible for a
49
branch to contain "lies" about its history or contents, but that
50
doesn't corrupt anything else.  It may confuse or mislead someone who
51
looks at the branch, but there is no substitute for human review
52
anyhow.
53
54
-------
55
56
The safest position may be to never rely on identifying content by
57
hash.  Rather, things which need a universally unique ID should get a
58
UUID instead.
59
60
This has a slight advantage that the id can be stored directly in the
61
object it refers to, when that's useful.
62
63
So a `Revision` holds a UUID for the `Inventory`.  
64
65
An inventory holds `InventoryEntry` objects, each with
66
67
* file-id
68
* filename (location in tree)
69
* type (file, dir, etc)
70
* text-id (uuid identifying the text)
71
* text-sha1
72
* text-length (for catching bugs)
73
* parent-file-id