| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
| |
"Why it holds" said a node takes a version only where it changes the one
the node reads, where the design has it take one it took, written again,
too.
|
| |
|
|
|
|
|
|
| |
What counts no more stayed, the text said, until a snapshot or a fork's
genesis left it out where it is or names a version before the current
one. But each holds a place's value alone: an envelope that isn't it,
behind another or a dropped key's, goes whatever version it names, and
a key added back brings back what it signed within the epoch only.
|
| |
|
|
|
|
|
|
| |
Since a party new in a version has its list named anew there, every party
has a list in every version from the one that let it in; the field said
lists were for the parties a version named one for. The example gave
makefu, its keys named anew at 7, none, and was no version. The field
says every party's, and the example gives makefu an empty list since 7.
|
| |
|
|
|
|
|
|
|
| |
A node follows no remake whose coordinator its party doesn't list, but a
node holding two decisions for its epoch followed neither, whatever they
were: a node taken off its list, by deciding a remake nobody follows,
still kept every node holding its decision from following an honest
remake of that epoch, until operators named one. A decision whose
coordinator isn't listed is none of the two.
|
| |
|
|
|
|
|
|
|
| |
A node refused the changes its operator had it drop. But any member
writes a version once it passed, and a change writing the dropped version
anew, by a lagging node by chance or a listed adversary at will, was none
of those: the node took it, as it changes the version the node reads, and
a node its operator was moving across a remake stayed. The node refuses
any change writing a version among those it dropped.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
The bound was judged at a change's history, as any rule is, so it
counted only the writers in that history: writers started side by side,
each at a history holding none of the others, were each within it, and a
party flooding on purpose passed it as often as it liked. Counted for a
party, it let one node gone wrong use up the room of all its party's
nodes for the rest of the epoch.
It goes. A party whose nodes start writers without end runs the others
out of memory: denial of service, which breaks no promise. Its writers
name its node, and the node's entry its party; a remake leaves an
epoch's writers behind, and the others fork without the party.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Read whole after the last series of changes, the design said a few things
that no longer fit what it says elsewhere:
* the example's version 7 named lass's keys anew at 4 but gave lass no
node list, where a party new in a version has its list named anew;
* a remake reaches no node that can't build it, nor one that isn't to
follow it (not a member at the base, its coordinator unlisted);
* the clause on whose word follows a decision sat after a colon meant for
the abort;
* what a node taken off its list could commit before a base ends with the
remake only where the base holds that list;
* a remake racing a listing leaves behind the nodes listed since the base,
not "the nodes that took it";
* "Why it holds" now names the coordinator's listing;
* attack 29: a node that lacks the base syncs it where a peer keeps it;
* forks come at stage 4, so before it no swarm starts afresh by one;
* a party drops a stolen key of its set itself; only one its version
names takes a version.
|
| |
|
|
|
|
|
|
|
|
| |
A version's `since`, where it doesn't name a party's keys or list anew,
is the `since` the version it changes gives the party, not that version's
number. A version written again is one a node took, and every node takes
it however often it is written. A forked writer's nodes are apart until
operators have one side drop what it took, or a fork. And a fork's
genesis leaves out continuum's own data, earlier epochs' metadata among
it, which a second clause said again.
|
| |
|
|
|
|
|
|
|
| |
A list had a `since` only for a party a version named one for. A party
taken out and let in again within an epoch had none unless the version
named it a list, so the list it left, signed with the keys it had then,
counted again at once, while its key set, its keys named anew, didn't.
A party new in a version has its list named anew there, empty where the
version gives it no nodes.
|
| |
|
|
|
|
|
|
|
| |
The bound on writers was a node's, and a party lists the nodes it
likes, a new list one envelope: the flood the bound was for cost a party
a list now and then. The bound is a party's: the writers its nodes start
in an epoch together, by the party each node's entry names, judged at a
change's history. Its figure leaves room for the writer an honest node
starts each time it starts, rejoins or drops, until a remake.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
A node taken off its party's list is a member at a history from before:
a remake's record, proposal and decision it writes there are taken, and
every node listed at its base follows, the snapshot listing the node
again until the list that took it off is replayed. "The next remake's
base" could be its own.
A node follows a remake only where the coordinator's party lists the
coordinator, as the node reads it when it switches; and what a node
taken off its list commits from a history before ends with a remake
whose base holds that list.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Continuum lets a member ack another remake of its epoch only once the
first is aborted, and in a federated swarm only the coordinator aborts:
a coordinator gone after proposing stalled every remake of its epoch.
Read the other way, as a remake left undecided binds nobody, members
moved on, and a coordinator back late decided beside the remake they
moved to: two decisions, from honest nodes alone.
Acks bind nobody: a member may ack another remake without the first
aborted, and an honest coordinator decides only while the acks it counts
are still for its remake.
|
| |
|
|
|
|
|
|
|
|
| |
A place's value is read among the envelopes a key their party holds
signed, but a remake's snapshot, a fork's genesis and the replay held
what the old document "read as", which continuum's merged view has as
the first of all the envelopes there, by `seq`: a builder reusing it
would keep a dropped key's envelope ahead of the party's, and nothing at
the place would count in the new epoch. Each holds, and replays to, the
place's value as the federation reads it.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
An operator may have its node drop what it took, and nothing said the
node refuses it afterwards: a peer that still holds it sends it again,
and the node, holding the votes that passed a dropped version, writes
it again itself, as a node writes any version it holds passed. So the
version a node left behind by a remake dropped came back before it
switched, and either side of keys that voted twice took again what it
dropped until it held the other side.
In that epoch a node refuses the changes its operator had it drop, from
whoever brings them, and writes no version among them again, until its
operator says otherwise.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A party's `true` on two proposals of one version was voting twice where
both passed, or neither was dead by all the votes. But a party's `true`
and `false` on one proposal both count, so one proposal can be passed
and dead at once: another party's late `true` beside its `false` made
every party that went on to the second proposal, once the first was
dead, look as if it had voted twice.
Two `true`s are voting twice only where neither proposal is dead by all
the votes. A death is never taken back, so an honest party's first
proposal is always dead by all of them; and where two versions changing
one were written, the party whose `true` and `false` made a proposal
passed and dead is named. "Why it holds" says the same.
|
| |
|
|
|
|
|
| |
Acks stay continuum's, by node, but in a federated swarm a node follows a
decision it can build without counting them, so what they decided was
nobody's to say. They tell an honest coordinator when to commit, so that
fewer nodes are away; they are no safeguard.
|
| |
|
|
|
|
|
|
| |
A remake that races a version leaves the nodes that took it in the old
epoch. Its operator moves such a node without a fork or an empty
document: it has the node drop the version it took, the base then holds
the version it reads, it switches, and the new epoch writes the version
again from the votes replayed, as any node writes one it holds passed.
|
| |
|
|
|
|
|
|
|
| |
Forks are what any group carries on by, and promise 3 rests on them;
remakes are there only for Automerge's sake. Forks come at stage 4 and
remakes at stage 5: until then a federated swarm that grows too large
starts afresh by a fork whose version names the same parties, keys and
rules. And a writer forked splits the nodes for good only until stage
2, where an operator may have its node drop what it took.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A version setting a party's keys or node list had the change writing it
write an envelope without a signature at the party's place, `prev`
empty and `seq` 0, and only that change, so that it was ahead of every
set before it: a kind of envelope of its own, a rule on who writes it,
the same written again with the version, and a genesis and a fork
writing them too.
Each version gives, for each party's keys and node list, the number of
the version that named them anew (`since`): its own where it does, else
that of the version it changes. A party's keys are those of its set
where the set names a version numbered at least that; otherwise those
the version names; its node list the same. Nothing is written for it,
the document holds no envelope without a signature, and what a key an
earlier version named writes at a history from before still counts no
more.
|
| |
|
|
|
|
|
|
|
|
|
| |
Where keys voted twice, or a writer forked, each node kept what it took,
and the sides came together only by a fork, a node with an empty
document, or its peers' word. An operator may have its node drop a
version, or any change, with everything built on it, as continuum drops
a change its node took and the rules refuse (`drop_taken`), and take
what the others hold: no peer's word moves a node, its operator's does.
So either side may join the other without what parts them, as the
design request asked, without a fork.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Which federation a swarm follows is in each version, not in its policy's
place; a fork's first version changes the one current at its heads, not
`null`; a place given back shows what its party signed before only where
nothing replaced it since; a remake's snapshot is a function of its base
and its record, so two remakes at one base are two epochs; honest nodes
agree once they sync but where keys voted twice, a coordinator decided
two remakes, or a writer forked; and what a node taken off its list
commits from a history before ends with the next remake's base, as no
later version orders it.
|
| |
|
|
|
|
|
|
|
| |
Every writer costs every node memory for as long as the epoch lasts, and
nothing bounded the writers a node starts: one listed node starting one
with every change runs every other out of memory before a remake could
be built, which the design called slow, not wrong. A federated swarm
bounds the writers a node starts in an epoch, judged at a change's
history as any rule is: a rule, and stage 1's.
|
| |
|
|
|
|
|
|
|
|
|
| |
A node reported a party that voted twice, judging each copy of its votes
by the history it was taken at, so that a vote committed again elsewhere
framed nobody; and the next remake left those histories out. The votes
alone tell it: a `true` and a `false` on one proposal, or a `true` on two
proposals of one version where both passed or neither is dead by all the
votes cast, as an honest party backs another only once the first is
dead. A tool over the votes the parties' clients keep reports it, later
too; a node reports the versions it refused.
|
| |
|
|
|
|
|
|
|
|
|
| |
A node took tokens from the nodes its party lists, and without a user
registry every user it lets in may do everything. A party's list is
written by a key its version names, or set by a version: so two thirds
of the parties, or a thief of that key, could list a node of their own
as the party's, issue a token under any name there, and have the party's
own nodes take it, against "the design rules nothing about what happens
on a party's own nodes". A node takes only its own tokens; a party logs
in to each of its nodes.
|
| |
|
|
|
|
|
|
| |
A node holding two versions that change the one it reads, having taken
neither, took neither until its operator named one. But where one comes
first it is taken anyway, and two come together only by how a sync
message was cut: the state, and the endpoint naming one, guarded
nothing. A node takes the first it judges.
|
| |
|
|
|
|
|
|
|
|
|
| |
A replay puts changes neither of which saw the other by their clocks.
A party's `true` committed again at a history before the nays that freed
it, its clock set between, came before those nays in the replay and was
refused, though the old epoch took it: a proposal that passed there might
not pass in the new one, and a party not finding its vote might vote
again. What the new epoch refuses is tried again once the rest is
replayed, and a party's client keeps what its party voted and votes by
that, never by what a node shows.
|
| |
|
|
|
|
|
|
|
|
| |
A proposal is taken whatever its text, and asked whether it is a version
only where it is written as one. One that isn't is voted on, never
passes, and dies only by nays: two thirds backing it bind the federation
to its version for good, and a client reading JSON otherwise than the
nodes backs what no node writes. A node shows whether a proposal is a
version by the federation it knows, and a party's client backs none
that isn't.
|
| |
|
|
|
|
|
|
| |
Without a user registry, every user a node lets in may do everything,
and `--protocol-writers` say who writes protocol data; what of that held
in a federated swarm the design didn't say. Whoever a node lets in may
have it commit what the rules take, envelopes and the node's own
protocol data; `--protocol-writers` names nobody.
|
| |
|
|
|
|
|
|
|
| |
Nothing bounded a version, a key set, a node list, a path or a value: a
version larger than a sync message carries would be written where it
passed and reach nobody, and nodes built with different bounds would
take different changes. The sizes each may have are the federation's
number's, each within what a sync message carries, and stage 1 sets
them.
|
| |
|
|
|
|
| |
A node follows a remake only where it is a member at the base: listed
there, or listed and registered, the design didn't say. Listed: a node
its party listed registers in the new epoch as it would have in the old.
|
| |
|
|
|
|
|
|
| |
Any member wrote a passed version; when, and which, the design didn't
say: every node that sees it passed, the one that took the last vote, or
one after a wait. A node writes it as soon as it holds it passed; the
same version written by several is one, as a node takes a version it
took written again.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
A federation's genesis change, judged by none, was its founding node's:
what else it held (values at places, votes) nobody checked, and the
swarm's name was the one its node drew or remembered, so two federations
started under one name took each other's envelopes. It is built byte for
byte from the first version alone, as a fork's genesis is from its
version and heads, and holds the policy, the swarm's name, the version
and what it sets, nothing else; the swarm is named by the version's
hash, and a node joining builds the genesis and checks it. The genesis
writes no topology: a federated swarm runs continuum's own.
|
| |
|
|
|
|
|
|
|
|
| |
A fork's genesis left out what a snapshot does, the federation's versions,
member entries and the swarm's name. It didn't say what of continuum's
other data it holds (remake records, acks, earlier epochs' metadata), nor
whether what counts no more is read by the fork's first version or the one
current at its heads: two founders deciding apart build two swarms where
one is meant. It holds none of continuum's own data, and reads what
counts no more by the version current at its heads.
|
| |
|
|
|
|
|
|
| |
A party writing first where another's envelope is (a place given to it,
an envelope a dropped key wrote) could name that one as replaced, `seq`
one more, as judging at a change's history needs, or start anew; two
builders deciding apart would refuse each other's envelopes. It names
that one.
|
| |
|
|
|
|
|
|
|
|
|
| |
A change wrote nothing where rules say who writes but `@` keys and the
maps on the way to them, never deleting or replacing one. It didn't say
that a map is made only on the way to an `@` key the same change writes:
one builder would take any member's maps in any party's namespace, left
empty for good, another wouldn't. Nor that a map is made only where none
is: a map made where one is, naming it as replaced, hides what it held,
votes under `_continuum/federation` among them. Both are said, for the
federation's own paths too.
|
| |
|
|
|
|
|
|
| |
Where a rule's `{party}` segment met a segment naming no party, one
builder would match it with a writer that is nobody, another wouldn't
match and let a later rule decide, and the two would judge a write at
`names/r/foo` apart where a `names/r/*` rule follows. It doesn't match,
which is the reading the wiki example needs.
|
| |
|
|
|
|
|
|
|
| |
Where a party's `true` and `false` on one proposal are both in one
history, the design didn't say whether it voted yes, no, both, or what
the place read as; one builder's node would take a version, or another
party's next `true`, where another builder's refused it. Both count:
then a history that holds more votes never takes back a pass or a
death.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A remake's decision was its coordinator's alone, but a member could abort
a remake that expired, as each node's clock said: the one place the
design waited on time. And a coordinator writing two decisions for one
epoch, showing each to some nodes, left honest nodes in two epochs for
good, as continuum's following of most peers, and rejoining sideways,
are gone.
The abort is the coordinator's too, and a remake left undecided binds
nobody. A node holding two decisions for its epoch follows neither, says
so, and leaves it to its operator; on its operator's word it builds the
remake named of the epoch before its own, from the old document it
keeps, and switches, what it wrote since replayed. Not promised and
Attacks say so.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A place's value was the first of all the envelopes it held, by version,
then `seq`, and it counted only while its key was its party's. So a key
the party had dropped, by a member committing at a history from before
the drop, wrote ahead of the party's envelope, replacing it, with a
higher `seq` each time the party wrote again: nothing at the place
counted until a version named the party's keys anew.
A place's value is the first of the envelopes there that a key their
party holds signed; judging, and the envelope a `prev` names, go by all
of them as before. And a party signs again what should stay after the
set that drops a key, not only before: what the dropped key writes at a
history from before can't name that as replaced, and counts no more, so
it takes nothing out.
|
| |
|
|
|
|
|
|
| |
Screening judges a batch at a frontier that every change touching what
the rules read moves, or a later change of the batch is judged at heads
it doesn't hold: stage 1 moves it for what the federation reads too. And
a node syncs with members only, but is one only once its own entry names
its party: stage 3 shows the party in the handshake before that.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
A node switches only where the version it reads is written at the base.
So a node that switches took no version written after it, and what it
replays names the version current at the base, or one before: the
replay of versions, each in one change with what it sets, after the
votes that passed it and whatever names the version before it, was
there for no node. It goes, and with it two `true`s cast at once,
replayed one after the other, undoing a version that passed.
The version a node reads is asked as it reads it when it switches: one
it takes while it builds the snapshot leaves it where it is, as one
taken before.
|
| |
|
|
|
|
|
|
|
|
|
| |
A node leaves its epoch only for one it built itself from a base it
holds; but continuum stops syncing the old epoch once it switched, so a
node away across a switch lacked some of the base and stayed until its
operator moved it, on most remakes where nodes come and go.
A node syncs the old document with nodes still on that epoch for as long
as its own setting keeps it, and a node that lacks some of the base
syncs it from them, each change judged there, and builds then.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
A node holding the old document left an epoch only for one it built
itself, but took a fork's heads' history, where it lacked some, on its
peers' word: a founder could name heads holding a change it sent nobody,
a version written without votes say, and the genesis would carry what
that change set. A node holding the old document takes a fork only where
it holds the heads' history, syncing what it lacks from the old swarm,
each change judged there; otherwise it says it can't check and leaves it
to its operator. Only a node with an empty document takes a fork on its
peers' word, as it takes an epoch.
|
| |
|
|
|
|
|
|
|
|
| |
Nothing is deleted within an epoch, and a snapshot left out only what
is or names a version before the current one, which no member entry
does: the entry of every node its party stopped listing, every node
reinstalled under a new ID among them, stayed for good. A snapshot
leaves out the entry of a node its party doesn't list at the base; a
node listed again registers again. A fork's genesis holds no member
entries already.
|
| |
|
|
|
|
|
|
|
|
| |
A node refuses a version that doesn't change the one it reads, valid at
its history or not. Continuum refuses as a whole a member whose own
change broke a rule, sending it nothing more; a version valid where it
was written breaks none, and its committer, often an honest member on
the other side of keys that voted twice, would stay cut off after the
operators settled. It isn't refused as a whole, and the node reports it
apart from changes against the rules.
|
| |
|
|
|
|
|
| |
A proposal was taken whatever its text, yet named by the hash of its
version: one builder took a proposal whose text hashes elsewhere, another
refused it, and their nodes parted. It is refused: the hash asks nothing
of what a version may say, so every node asks it alike.
|
| |
|
|
|
|
|
|
|
|
|
| |
A node took a version only where it changed the one it read, or was
that one written again. A member that was away writes a passed version
itself, at a history where the version before is current, and the
others took the next version meanwhile: its change writes neither the
version they read nor one changing it, so they refused it and everything
the member built on it, an honest member split off.
A node takes a version written again wherever it is one the node took.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
What the design leaned on that didn't hold, and what it can do without:
* A vote counts as cast wherever its version is current, so the keys an
earlier version named could pass another proposal on it at a history
from before the version that followed, raise its number alone, and be
read by every node. A node takes a version only where it changes the
one it reads; another it refuses, reports, and leaves to its operator.
Two that change one are written only where keys voted twice: each node
keeps the one it took, and the parties fork. The lower hash goes.
* A node left its epoch on its peers' word (catching up), or for a
remake at any base its coordinator named. It leaves an epoch only for
a compaction it built itself, from a base it holds, a member there,
the version it reads written there; the decision is the coordinator's
alone, and how long the old document is kept the node's own setting.
The check that a remake reads as the old document goes, with remakes
of other kinds.
* Housekeeping within an epoch shrank nothing (Automerge keeps every
operation), and an emptied place read whatever anyone put back.
Nothing is deleted in an epoch but by a party's own `null`; a
snapshot and a fork's genesis leave out what counts no more and is or
names a version before the current one. Rule 7, putting back and the
replay's deletions go.
* A key could hold an envelope and a map at once. A place's value is in
its `@` key, which never holds a map, so no key is both; `**` stays.
* A key set dropping the key that signed it counted no more by the rule
for values. Sets are checked only where written, and signed, as node
lists are, only by keys the version names, so a party revokes a key
of its set by itself. What a version sets for a party (keys, and now
node lists, as the first version and a fork's do, so that a group
without listed nodes can fork) is one envelope, `prev` empty, `seq` 0.
* The federation's envelope is `cv3.`, so that swarms that aren't
federated read nothing new; neighbours are kept partly by party; the
krebs users who will vote are the parties, the others in a group.
* The share is two thirds, the field gone; `went/` goes.
And what two builders would decide apart: a version's bytes its JCS
encoding, validity asked where it is written as a version, the rule
grammar's edges, keys holding `/`, the double-vote report against a
vote committed again elsewhere, one list of the protocol data a
federated swarm takes, the Automerge version a rule, releases working
together, the swarm key, a yes that can't be taken back, what stage 1
covers and that its swarms are for tests.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Nothing made a federation's first node a member, so nothing after its
genesis change was ever taken. The genesis change holds its node's
party's node list, naming that node, and the other parties' lists come
in through it. A list counts as its party's, whichever of its keys
signed it, until the party writes another: no change of keys leaves a
swarm or a fork without members, and signing the list again with the
keys goes.
A fork's genesis is said to be a function of its first version and the
heads alone. Closed is at, within and above, a member's entry too, and
the swarm's topology is closed with the rest. Smaller: keys differ by
keys; a party changing keys adds the new ones, signs again with one,
then drops the old; the keys a version sets name the version being
written; rule 6 is read at the deleting change's history; a party may
keep the keys its version names offline. Rows for the genesis list,
stranded lists and a stolen voting key.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A vote and a proposal now count by the keys their version names for
their party: every version names every party's keys, and the parties
are the names of those keys. So a vote is checked from itself and its
version alone, and a remake replays it as cast in whatever order key
sets of its party go around it; replaying votes before key sets, which
no order could meet where two parties' key sets and votes cross, goes,
with the rule treating two key sets at once as voting twice. A party's
own key change reaches its votes with the next version naming its keys.
A federation's genesis change is its swarm's first, so it starts empty:
content before it would be a change judged by none. A fork's genesis is
built byte for byte, so one version and one set of heads make one
swarm. A member's entry is its member's only, removed whole; logouts,
sessions and forgotten nodes are closed, and a write at the policy's
place refused. Rule 7 waits for what was written beside a `null` too; a
`null` housekeeping deleted after a base is back for an epoch. A party
signs its node list again with its keys. "Why it holds" says what a
remake keeps. Rows for what this answers.
|