mirror of
https://github.com/ilyakooo0/urbit.git
synced 2024-11-28 19:55:53 +03:00
added clay architecture doc
This commit is contained in:
parent
646fa7cf14
commit
0bfa034edb
415
pub/doc/arvo/clay/architecture.md
Normal file
415
pub/doc/arvo/clay/architecture.md
Normal file
@ -0,0 +1,415 @@
|
||||
# clay
|
||||
|
||||
## high-level
|
||||
|
||||
clay is the primary filesystem for the arvo operating system,
|
||||
which is the core of an urbit. The architecture of clay is
|
||||
intrinsically connected with arvo, but we assume no knowledge of
|
||||
either arvo or urbit. We will point out only those features of
|
||||
arvo that are necessary for an understanding of clay, and we will
|
||||
do so only when they arise.
|
||||
|
||||
The first relevant feature of arvo is that it is a deterministic
|
||||
system where input and output are defined as a series of events
|
||||
and effects. The state of arvo is simply a function of its event
|
||||
log. None of the effects from an event are emitted until the
|
||||
event is entered in the log and persisted, either to disk or
|
||||
another trusted source of persistence, such as a Kafka cluster.
|
||||
Consequently, arvo is a single-level store: everything in its
|
||||
state is persistent.
|
||||
|
||||
In a more traditional OS, everything in RAM can be erased at any
|
||||
time by power failure, and is always erased on reboot. Thus, a
|
||||
primary purpose of a filesystem is to ensure files persist across
|
||||
power failures and reboots. In arvo, both power failures and
|
||||
reboots are special cases of suspending computation, which is
|
||||
done safely since our event log is already persistent. Therefore,
|
||||
clay is not needed in arvo for persistence. Why, then, do we have a
|
||||
filesystem? There are two answers to this question.
|
||||
|
||||
First, clay provides a filesystem tree, which is a convenient
|
||||
user interface for some applications. Unix has the useful concept
|
||||
of virtual filesystems, which are used for everything from direct
|
||||
access to devices, to random number generators, to the /proc
|
||||
tree. It is easy and intuitive to read from and write to a
|
||||
filesystem tree.
|
||||
|
||||
Second, clay has a distributed revision control system baked into
|
||||
it. Traditional filesystems are not revision controlled, so
|
||||
userspace software -- such as git -- is written on top of them to
|
||||
do so. clay natively provides the same functionality as modern
|
||||
DVCSes, and more.
|
||||
|
||||
clay has two other unique properties that we'll cover later on:
|
||||
it supports typed data and is referentially transparent.
|
||||
|
||||
### Revision Control
|
||||
|
||||
Every urbit has one or more "desks", which are independently
|
||||
revision-controlled branches. Each desk contains its own mark
|
||||
definitions, apps, doc, and so forth.
|
||||
|
||||
Traditionally, an urbit has at least a base and a home desk. The
|
||||
base desk has all the system software from the distribution. the
|
||||
home desk is a fork of base with all the stuff specific to the
|
||||
user of the urbit.
|
||||
|
||||
A desk is a series of numbered commits, the most recent of which
|
||||
represents the current state of the desk. A commit is composed of
|
||||
(1) an absolute time when it was created, (2) a list of zero or
|
||||
more parents, and (3) a map from paths to data.
|
||||
|
||||
Most commits have exactly one parent, but the initial commit on a
|
||||
desk may have zero parents, and merge commits have more than one
|
||||
parent.
|
||||
|
||||
The non-meta data is stored in the map of paths to data. It's
|
||||
worth noting that no constraints are put on this map, so, for
|
||||
example, both /a/b and /a/b/c could have data. This is impossible
|
||||
in a traditional Unix filesystem since it means that /a/b is both
|
||||
a file and a directory. Conventionally, the final element in the
|
||||
path is its mark -- much like a filename extension in Unix. Thus,
|
||||
/doc/readme.md in Unix is stored as /doc/readme/md in urbit.
|
||||
|
||||
The data is not stored directly in the map; rather, a hash of the
|
||||
data is stored, and we maintain a master blob store. Thus, if the
|
||||
same data is referred to in multiple commits (as, for example,
|
||||
when a file doesn't change between commits), only the hash is
|
||||
duplicated.
|
||||
|
||||
In the master blob store, we either store the data directly, or
|
||||
else we store a diff against another blob. The hash is dependent
|
||||
only on the data within and not on whether or not it's stored
|
||||
directly, so we may on occasion rearrange the contents of the
|
||||
blob store for performance reasons.
|
||||
|
||||
Recall that a desk is a series of numbered commits. Not every
|
||||
commit in a desk must be numbered. For example, if the base desk
|
||||
has had 50 commits since home was forked from it, then a merge
|
||||
from base to home will only add a single revision number to home,
|
||||
although the full commit history will be accessible by traversing
|
||||
the parentage of the individual commits.
|
||||
|
||||
We do guarantee that the first commit is numbered 1, commits are
|
||||
numbered consecutively after that (i.e. there are no "holes"),
|
||||
the topmost commit is always numbered, and every numbered commit
|
||||
is an ancestor of every later numbered commit.
|
||||
|
||||
There are three ways to refer to particular commits in the
|
||||
revision history. Firstly, one can use the revision number.
|
||||
Secondly, one can use any absolute time between the one numbered
|
||||
commit and the next (inclusive of the first, exclusive of the
|
||||
second). Thirdly, every desk has a map of labels to revision
|
||||
numbers. These labels may be used to refer to specific commits.
|
||||
|
||||
Additionally, clay is a global filesystem, so data on other urbit
|
||||
is easily accessible the same way as data on our local urbit. In
|
||||
general, the path to a particular revision of a desk is
|
||||
/~urbit-name/desk-name/revision. Thus, to get /try/readme/md
|
||||
from revision 5 of the home desk on ~sampel-sipnym, we refer to
|
||||
/~sampel-sipnym/home/5/try/readme/md. Clay's namespace is thus
|
||||
global and referentially transparent.
|
||||
|
||||
XXX reactivity here?
|
||||
|
||||
### A Typed Filesystem
|
||||
|
||||
Since clay is a general filesystem for storing data of arbitrary
|
||||
types, in order to revision control correctly it needs to be
|
||||
aware of types all the way through. Traditional revision control
|
||||
does an excellent job of handling source code, so for source code
|
||||
we act very similar to traditional revision control. The
|
||||
challenge is to handle other data similarly well.
|
||||
|
||||
For example, modern VCSs generally support "binary files", which
|
||||
are files for which the standard textual diffing, patching, and
|
||||
merging algorithms are not helpful. A "diff" of two binary files
|
||||
is just a pair of the files, "patching" this diff is just
|
||||
replacing the old file with the new one, and "merging"
|
||||
non-identical diffs is always a conflict, which can't even be
|
||||
helpfully annotated. Without knowing anything about the structure
|
||||
of a blob of data, this is the best we can do.
|
||||
|
||||
Often, though, "binary" files have some internal structure, and
|
||||
it is possible to create diff, patch, and merge algorithms that
|
||||
take advantage of this structure. An image may be the result of a
|
||||
base image with some set of operations applied. With algorithms
|
||||
aware of this set of operations, not only can revision control
|
||||
software save space by not having to save every revision of the
|
||||
image individually, these transformations can be made on parallel
|
||||
branches and merged at will.
|
||||
|
||||
Suppose Alice is tasked with touching up a picture, improving the
|
||||
color balance, adjusting the contrast, and so forth, while Bob
|
||||
has the job of cropping the picture to fit where it's needed and
|
||||
adding textual overlay. Without type-aware revision control,
|
||||
these changes must be made serially, requiring Alice and Bob to
|
||||
explicitly coordinate their efforts. With type-aware revision
|
||||
control, these operations may be performed in parallel, and then
|
||||
the two changesets can be merged programmatically.
|
||||
|
||||
Of course, even some kinds of text files may be better served by
|
||||
diff, patch, and merge algorithms aware of the structure of the
|
||||
files. Consider a file containing a pretty-printed JSON object.
|
||||
Small changes in the JSON object may result in rather significant
|
||||
changes in how the object is pretty-printed (for example, by
|
||||
addding an indentation level, splitting a single line into
|
||||
multiple lines).
|
||||
|
||||
A text file wrapped at 80 columns also reacts suboptimally with
|
||||
unadorned Hunt-McIlroy diffs. A single word inserted in a
|
||||
paragraph may push the final word or two of the line onto the
|
||||
next line, and the entire rest of the paragraph may be flagged as
|
||||
a change. Two diffs consisting of a single added word to
|
||||
different sentences may be flagged as a conflict. In general,
|
||||
prose should be diffed by sentence, not by line.
|
||||
|
||||
As far as the author is aware, clay is the first generalized,
|
||||
type-aware revision control system. We'll go into the workings
|
||||
of this system in some detail.
|
||||
|
||||
### Marks
|
||||
|
||||
Central to a typed filesystem is the idea of types. In clay, we
|
||||
call these "marks". A mark is a file that defines a type,
|
||||
conversion routines to and from the mark, and diff, patch, and
|
||||
merge routines.
|
||||
|
||||
For example, a `%txt` mark may be a list of lines of text, and it
|
||||
may include conversions to `%mime` to allow it to be serialized
|
||||
and sent to a browswer or to the unix filesystem. It will also
|
||||
include Hunt-McIlroy diff, patch, and merge algorithms.
|
||||
|
||||
A `%json` mark would be defined as a json object in the code, and
|
||||
it would have a parser to convert from `%txt` and a printer to
|
||||
convert back to `%txt`. The diff, patch, and merge algorithms are
|
||||
fairly straightforward for json, though they're very different
|
||||
from the text ones.
|
||||
|
||||
More formally, a mark is a core with three arms, `++grab`,
|
||||
`++grow`, and `++grad`. In `++grab` is a series of functions to
|
||||
convert from other marks to the given mark. In `++grow` is a
|
||||
series of functions to convert from the given mark to other
|
||||
marks. In `++grad` is `++diff`, `++pact`, `++join`, and `++mash`.
|
||||
|
||||
The types are as follows, in an informal pseudocode:
|
||||
|
||||
++ grab:
|
||||
++ mime: <mime> -> <mark-type>
|
||||
++ txt: <txt> -> <mark-type>
|
||||
...
|
||||
++ grow:
|
||||
++ mime: <mark-type> -> <mime>
|
||||
++ txt: <mark-type> -> <txt>
|
||||
...
|
||||
++ grad
|
||||
++ diff: (<mark-type>, <mark-type>) -> <diff-type>
|
||||
++ pact: (<mark-type>, <diff-type>) -> <mark-type>
|
||||
++ join: (<diff-type>, <diff-type>) -> <diff-type> or NULL
|
||||
++ mash: (<diff-type>, <diff-type>) -> <diff-type>
|
||||
|
||||
These types are basically what you would expect. Not every mark
|
||||
has each of these functions defined -- all of them are optional
|
||||
in the general case.
|
||||
|
||||
In general, for a particular mark, the `++grab` and `++grow` entries
|
||||
(if they exist) should be inverses of each other.
|
||||
|
||||
In `++grad`, `++diff` takes two instances of a mark and produces
|
||||
a diff of them. `++pact` takes an instance of a mark and patches
|
||||
it with the given diff. `++join` takes two diffs and attempts to
|
||||
merge them into a single diff. If there are conflicts, it
|
||||
produces null. `++mash` takes two diffs and forces a merge,
|
||||
annotating any conflicts.
|
||||
|
||||
In general, if `++diff` called with A and B produces diff D, then
|
||||
`++pact` called with A and D should produce B. Also, if `++join`
|
||||
of two diffs does not produce null, then `++mash` of the same
|
||||
diffs should produce the same result.
|
||||
|
||||
Alternately, instead of `++diff`, `++pact`, `++join`, and
|
||||
`++mash`, a mark can provide the same functionality by defining
|
||||
`++sted` to be the name of another mark to which we wish to
|
||||
delegate the revision control responsibilities. Then, before
|
||||
running any of those functions, clay will convert to the other
|
||||
mark, and convert back afterward. For example, the `%hoon` mark
|
||||
is revision-controlled in the same way as `%txt`, so its `++grad`
|
||||
is simply `++sted %txt`. Of course, `++txt` must be defined in
|
||||
`++grow` and `++grab` as well.
|
||||
|
||||
Every file in clay has a mark, and that mark must have a
|
||||
fully-functioning `++grad`. Marks are used for more than just
|
||||
clay, and other marks don't need a `++grad`, but if a piece of
|
||||
data is to be saved to clay, we must know how to revision-control
|
||||
it.
|
||||
|
||||
Additionally, if a file is to be synced out to unix, then it must
|
||||
have conversion routines to and from the `%mime` mark.
|
||||
|
||||
##Using clay
|
||||
|
||||
### Reading and Subscribing
|
||||
|
||||
When reading from Clay, there are three types of requests. A
|
||||
`%sing` request asks for data at single revsion. A `%next`
|
||||
request asks to be notified the next time there's a change to
|
||||
given file. A `%many` request asks to be notified on every
|
||||
change in a desk for a range of changes.
|
||||
|
||||
For `%sing` and `%next`, there are generally three things to be
|
||||
queried. A `%u` request simply checks for the existence of a
|
||||
file at a path. A `%x` request gets the data in the file at a
|
||||
path. A `%y` request gets a hash of the data in the file at the
|
||||
path combined with all its children and their data. Thus, `%y`
|
||||
of a node changes if it or any of its children change.
|
||||
|
||||
A `%sing` request is fulfilled immediately if possible. If the
|
||||
requested revision is in the future, or is on another ship for
|
||||
which we don't have the result cached, we don't respond
|
||||
immediately. If the requested revision is in the future, we wait
|
||||
until the revision happens before we respond to the request. If
|
||||
the request is for data on another ship, we pass on the request
|
||||
to the other ship. In general, Clay subscriptions, like most
|
||||
things in Urbit, aren't guaranteed to return immediately.
|
||||
They'll return when they can, and they'll do so in a
|
||||
referentially transparent manner.
|
||||
|
||||
A `%next` request checks query at the given revision, and it
|
||||
produces the result of the query the next time it changes, along
|
||||
with the revsion number when it changes. Thus, a `%next` of a
|
||||
`%u` is triggered when a file is added or deleted, a `%next of a
|
||||
`%x` is triggered when a file is added, deleted, or changed, and
|
||||
a `%next` of a `%y` is triggered when a file or any of its
|
||||
children is added, deleted, or changed.
|
||||
|
||||
A `%many` request is triggered every time the given desk has a
|
||||
new revision. Unlike a `%next`, a `%many` has both a start and
|
||||
an end revsion, after which it stops returning. For `%next`, a
|
||||
single change is reported, and if the caller wishes to hear of
|
||||
the next change, it must resubscribe. For `%many`, every revsion
|
||||
from the start to the end triggers a response. Since a `%many`
|
||||
request doesn't ask for any particular data, there aren't `%u`,
|
||||
`%x`, and `%y` versions for it.
|
||||
|
||||
### Unix sync
|
||||
|
||||
One of the primary functions of clay is as a convenient user
|
||||
interface. While tools exist to use clay from within urbit, it's
|
||||
often useful to be able to treat clay like any other filesystem
|
||||
from the Unix perspective -- to "mount" it, as it were.
|
||||
|
||||
From urbit, you can run `|mount /path/to/directory %mount-point`,
|
||||
and this will mount the given clay directory to the mount-point
|
||||
directory in Unix. Every file is converted to `%mime` before it's
|
||||
written to Unix, and converted back when read from Unix. The
|
||||
entire directory is watched (a la Dropbox), and every change is
|
||||
auto-committed to clay.
|
||||
|
||||
### Merging
|
||||
|
||||
Merging is a fundamental operation for a distributed revision
|
||||
control system. At their root, clay's merges are similar to
|
||||
git's, but with some additions to accomodate typed data. There
|
||||
are seven different merge strategies.
|
||||
|
||||
Throughout our discussion, we'll say that the merge is from
|
||||
Alice's desk to Bob's. Recall that a commit is a date (for all
|
||||
new commits this will be the current date), a list of parents,
|
||||
and the data itself.
|
||||
|
||||
A `%init` merge should be used iff it's the first commit to a
|
||||
desk. The head of Alice's desk is used as the number 1 commit to
|
||||
Bob's desk. Obviously, the ancestry remains intact through
|
||||
traversing the parentage of the commit even though previous
|
||||
commits are not numbered for Bob's desk.
|
||||
|
||||
A `%this` merge means to keep what's in Bob's desk, but join the
|
||||
ancestry. Thus, the new commit has the head of each desk as
|
||||
parents, but the data is exactly what's in Bob's desk. For those
|
||||
following along in git, this is the 'ours' merge strategy, not
|
||||
the '--ours' option to the 'recursive' merge strategy. In other
|
||||
words, even if Alice makes a change that does not conflict with
|
||||
Bob, we throw it away. It's Bob's way or the highway.
|
||||
|
||||
A `%that` merge means to take what's in Alice's desk, but join
|
||||
the ancestry. This is the reverse of `%this`.
|
||||
|
||||
A `%fine` merge is a "fast-forward" merge. This succeeds iff one
|
||||
head is in the ancestry of the other. In this case, we use the
|
||||
descendant as our new head.
|
||||
|
||||
For `%meet`, `%mate`, and `%meld` merges, we first find the most
|
||||
recent common ancestor to use as our merge base. If we have no
|
||||
common ancestors, then we fail. If we have more than one most
|
||||
recent common ancestor, then we have a criss-cross situation,
|
||||
which should be handled delicately. At present, we delicately
|
||||
throw up our hands and give up, but something akin to git's
|
||||
'recursive' strategy should be implemented in the future.
|
||||
|
||||
There's a functional inclusion ordering on `%fine`, `%meet`,
|
||||
`%mate`, and `%meld` such that if an earlier strategy would have
|
||||
succeeded, then every later strategy will produce the same
|
||||
result. Put another way, every earlier strategy is the same as
|
||||
every later strategy except with a restricted domain.
|
||||
|
||||
A `%meet` merge only succeeds if the changes from the merge base
|
||||
to Alice's head (hereafter, "Alice's changes") are in different
|
||||
files than Bob's changes. In this case, the parents are both
|
||||
Alice's and Bob's heads, and the data is the merge base plus
|
||||
Alice's changed files plus Bob's changed files.
|
||||
|
||||
A `%mate` merge attempts to merge changes to the same file when
|
||||
both Alice and bob change it. If the merge is clean, we use it;
|
||||
otherwise, we fail. A merge between different types of changes --
|
||||
for example, deleting a file vs changing it -- is always a
|
||||
conflict. If we succeed, the parents are both Alice's and Bob's
|
||||
heads, and the data is the merge base plus Alice's changed files
|
||||
plus Bob's changed files plus the merged files.
|
||||
|
||||
A `%meld` merge will succeed even if there are conflicts. If
|
||||
there are conflicts in a file, then we use the merge base's
|
||||
version of that file, and we produce a set of files with
|
||||
conflicts. The parents are both Alice's and Bob's heads, and the
|
||||
data is the merge base plus Alice's changed files plus Bob's
|
||||
changed files plus the successfully merged files plus the merge
|
||||
base's version of the conflicting files.
|
||||
|
||||
That's the extent of the merge options in clay proper. In
|
||||
userspace there's a final option `%auto`, which is the most
|
||||
common. `%auto` checks to see if Bob's desk exists, and if it
|
||||
doesn't we use a `%init` merge. Otherwise, we progressively try
|
||||
`%fine`, `%meet`, and `%mate` until one succeeds.
|
||||
|
||||
If none succeed, we merge Bob's desk into a scratch desk. Then,
|
||||
we merge Alice's desk into the scratch desk with the `%meld`
|
||||
option to force the merge. For each file in the produced set of
|
||||
conflicting files, we call the `++mash` function for the
|
||||
appropriate mark, which annotates the conflicts if we know how.
|
||||
|
||||
Finally, we display a message to the user informing them of the
|
||||
scratch desk's existence, which files have annotated conflicts,
|
||||
and which files have unannotated conflicts. When the user has
|
||||
resolved the conflicts, they can merge the scratch desk back into
|
||||
Bob's desk. This will be a `%fine` merge since Bob's head is in
|
||||
the ancestry of the scratch desk.
|
||||
|
||||
### Autosync
|
||||
|
||||
Tracking and staying in sync with another desk is another
|
||||
fundamental operation. We call this "autosync". This doesn't mean
|
||||
simply mirroring a desk, since that wouldn't allow local changes.
|
||||
We simply want to apply changes as they are made upstream, as
|
||||
long as there are no conflicts with local changes.
|
||||
|
||||
This is implemented by watching the other desk, and, when it has
|
||||
changes, merging these changes into our desk with the usual merge
|
||||
strategies.
|
||||
|
||||
Note that it's quite reasonable for two desks to be autosynced to
|
||||
each other. This results in any change on one desk being mirrored
|
||||
to the other and vice versa.
|
||||
|
||||
Additionally, it's fine to set up an autosync even if one desk,
|
||||
the other desk, or both desks do not exist. The sync will be
|
||||
activated when the upstream desk comes into existence and will
|
||||
create the downstream desk if needed.
|
Loading…
Reference in New Issue
Block a user