1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
|
-*- indented-text -*-
Id: TODO,v 1.29 2001/03/18 03:12:20 mbp Exp
* Most urgent: do rolling checksums rather than from-scratch if
possible.
* Don't use the rs_buffers_t structure.
There's something confusing about the existence of this structure.
In part it may eb the name. I think people expect that it will be
something that behaves like a FILE* or C++ stream, and it really
does not. Also, the structure does not behave as an object: it's
really just a shorthand for passing values in to the encoding
routines, and so does not have a lot of identity of its own.
An alternative might be
result = rs_job_iter(job,
in_buf, &in_len, in_is_ending,
out_buf, &out_len);
where we update the length parameters on return to show how much we
really consumed.
One technicality here will be to restructure the code so that the
input buffers are passed down to the scoop/tube functions that need
them, which are relatively deeply embedded. I guess we could just
stick them into the job structure, which is becoming a kind of
catch-all "environment" for poor C programmers.
* Meta-programming
* Plot lengths of each function
* Some kind of statistics on delta each day
* Encoding format
* Include a version in the signature and difference fields
* Remember to update them if we ever ship a buggy version (nah!) so
that other parties can know not to trust the encoded data.
* abstract encoding
In fact, we can vary on several different variables:
* what signature format are we using
* what command protocol are we using
* what search algorithm are we using?
* what implementation version are we?
Some are more likely to change than others. We need a chart
showing which source files depend on which variable.
* Error handling
* What happens if the user terminates the request?
* Do HTTP CONNECT
* This might be a nice place to use select!
* Encoding implementation
* Join up copy commands through the copyq if this is not done yet.
* Join up signature commands
* Encoding algorithm
* Self-referential copy commands
Suppose we have a file with repeating blocks. The gdiff format
allows for COPY commands to extend into the *output* file so that
they can easily point this out. By doing this, they get
compression as well as differencing.
It'd be pretty simple to implement this, I think: as we produce
output, we'd also generate checksums (using the search block
size), and add them to the sum set. Then matches will fall out
automatically, although we might have to specially allow for
short blocks.
However, I don't see many files which have repeated 1kB chunks,
so I don't know if it would be worthwhile.
* Extended files
Suppose the new file just has data added to the end. At the
moment, we'll match everything but the last block of the old
file. It won't match, because at the moment the search block
size is only reduced at the end of the *new* file. This is a
little inefficient, because ideally we'd know to look for the
last block using the shortened length.
This is a little hard to implement, though perhaps not
impossible. The current rolling search algorithm can only look
for one block size at any time. Can we do better? Can we look
for all block lengths that could match anything?
Remember also that at the moment we don't send the block length
in the signature; it's implied by the length of the new block
that it matches. This is kind of cute, and importantly helps
reduce the length of the signature.
* State-machine searching
Building a state machine from a regular expression is a brilliant
idea. (I think `The Practice of Programming' walks through the
construction of this at a fairly simple level.)
In particular, we can search for any of a large number of
alternatives in a very efficient way, with much less effort than
it would take to search for each the hard way. Remember also the
string-searching algorithms and how much time they can take.
I wonder if we can use similar principles here rather than the
current simple rolling-sum mechanism? Could it let us match
variable-length signatures?
* Cross-file matches
If the downstream server had many similar URLs, it might be nice
if it could draw on all of them as a basis. At the moment
there's no way to express this, and I think the work of sending
up signatures for all of them may be too hard.
Better just to make sure we choose the best basis if there is
none present. Perhaps this needs to weigh several factors.
One factor might be that larger files are better because they're
more likely to have a match. I'm not sure if that's very strong,
because they'll just bloat the request. Another is that more
recent files might be more useful.
* Support gzip compression of the difference stream. Does this
belong here, or should it be in the client and librsync just have
an interface that lets it cleanly plug in?
* Licensing
* Will the GNU Lesser GPL work? Specifically, will it be a problem
in distributing this with Mozilla or Apache?
* Checksums
* Do we really need to require that signatures arrive after the
data they describe? Does it make sense in HTTP to resume an
interrupted transfer?
I hope we can do this. If we can't, however, then we should
relax this constraint and allow signatures to arrive before the
data they describe. (Really? Do we care?)
* Allow variable-length checksums in the signature; the signature
will have to describe the length of the sums and we must compare
them taking this into account.
* Testing
* test broken pipes
* Test files >2GB, >4GB. Presumably these must be done in streams
so that the disk requirements to run the test suite are not too
ridiculous. I wonder if it will take too long to run these
tests? Probably, but perhaps we can afford to run just one
carefully-chosen test.
* Use slprintf not strnprintf, etc.
* Long files
* How do we handle the large signatures required to support large
files? In particular, how do we choose an appropriate block size
when the length is unknown? Perhaps we should allow a way for
the signature to scale up as it grows.
* What do we need to do to compile in support for this?
* On GNU, defining _LARGEFILE_SOURCE as we now do should be
sufficient.
* SCO and similar things on 32-bit platforms may be more
difficult. Some SCO systems have no 64-bit types at all, so
there we will have to do without.
* On larger Unix platforms we hope that large file support will
be the default.
* Perhaps make extracted signatures still be wrapped in commands.
What would this lead to?
* We'd know how much signature data we expect to read, rather than
requiring it to be terminated by the caller.
* Selective trace of particular areas of the library.
|
