| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
| |
Nodes push changes to their neighbours but don't relay them directly.
In a partial mesh, nodes can learn about missing changes via keepalive
messages and can pull them from their neighbours.
|
| | |
|
| | |
|
| |
|
|
|
| |
To prevent stale hello replies to cause joins and thus stale sessions,
each run gets a random ID, which used as nonce in the handshake.
|
| | |
|
| |
|
|
|
| |
Don't store encoded messages but retry using a newly generated one.
That way any accumulated changes can be delivered at once.
|
| |
|
|
|
| |
Wait 2*keepalive but at least 20 s to trigger resync so that lost
messages can be resent first.
|
| | |
|
| |
|
|
|
|
| |
The stall breaker was superseded by regular re-handshaking Re-handshaking also
gets triggered, when a node needs some of the heads its peer believes it has
already sent. In that case the peer now resends what's needed.
|
| |
|
|
|
|
|
|
|
|
| |
Ready nodes will leave the first message to the pear that isn't ready.
Hello and hello reply now carry ready, so readiness is known from the
very first contact, before any sync. Together with the sync messages
and keepalives, reports_ready is always current.
After epoch switch, pings go to everyone except ready members already
on the new epoch, these learn from the lagging peer's opening.
|
| | |
|
| |
|
|
|
| |
The handshake resets the session anyway, and the sender's reply timeout
resends what still matters.
|
| |
|
|
|
| |
Instead of arbitrarily limiting the number to 16, allow filling
keepalive messages as long as they don't cause mycelium to chunk it.
|
| |
|
|
| |
This allows changing sources without the fixed 30 s stall duration.
|
| |
|
|
|
|
|
| |
Add ready and syncs_with_you flags as well as actual heads, so the peer
- ready: same meaning as in the sync message
- syncs_with_you: sender counts recipient as member and in same epoch
- heads: allows to tell who's behind or if both sides have changes
|
| |
|
|
|
|
| |
Sync messages now carries a ready flag, telling whether the sending peer
itself is joined or is itself still (re-)joining. In the latter case,
the peer won't be used as sync source.
|
| | |
|
| |
|
|
|
| |
This prevents a fall back to the 3 s grace period if a reset happens
between the epoch switch and our opening went out.
|
| | |
|
| | |
|
| |
|
|
| |
SyncSession holds all fields that get reset by reset_sync()
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
| |
If both sides open, the openings cross, and an empty node's answer to
the other opening (empty heads) makes the other side send the whole
document a second time.
Now after a sync reset, only the side with the smaller public key opens
the sync; the other answers it, or opens itself if nothing arrives.
|
| |
|
|
|
|
|
|
|
|
|
| |
When an sync messages arrives and automerge has nothing to say about it,
then no reply is generated. With c9279e4 it was assumed that every sync
message gets an answer, so missing replies caused unnecessary resends.
Now a reply is only awaited if the peer's last reported heads differ
from ours. If the peer's heads become equal through a third peer without
replying, the waiting will also abort if a ping or pong showss the same
heads digest on both sides.
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
| |
GET /store?heads=true and GET /store/{key}?heads=true add the header
x-continuum-heads (sorted, comma-separated hex). The heads come from
the same moment as the read.
- It's opt-in because reading the heads takes the write lock.
Reads without the flag don't change.
- Stale reads during a rejoin never carry heads.
|
| | |
|
| | |
|
| | |
|
| |
|
|
| |
This prevents an immediate ping during handshake.
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
