- alertmanager_cluster_alive_messages_total, total number of alive
messages received.
- alertmanager_cluster_peer_info, a constant metric labeled by peer name.
- alertmanager_cluster_pings_seconds, histogram of latencies for ping
messages.
Signed-off-by: Simon Pasquier <spasquie@redhat.com>
Adds a job which runs periodically and refreshes cluster.peer dns records.
The problem is that when you restart all of the alertmanager instances in an environment like Kubernetes, DNS may contain old alertmanager instance IPs, but on startup (when Join() happens) none of the new instance IPs. As at the start DNS is not empty resolvePeers waitIfEmpty=true, will return and "islands" of 1 alertmanager instances will form.
Signed-off-by: Povilas Versockas <p.versockas@gmail.com>
Alertmanager is exiting with a non-zero exit code if the initial cluster
join fails. This behavior could be not wanted because:
- As Alertmanager is a critical component with an at-least-once
guarantee, failing on joining the cluster is unnecessary as
Alertmanager still functions by itself.
- In an environment like Kubernetes discovering peers via DNS, peers
might roll out one-by-one, leaving the DNS entries unpopulated for the
first peer of a set. Failing on initial join prevents a roll-out.
Instead of failing on the initial join this patch only logs the failure.
The cluster can be later joined via the `handleReconnect`.
This is a regression introduced in PR #1456 [1].
[1] https://github.com/prometheus/alertmanager/pull/1456
Signed-off-by: Max Leonard Inden <IndenML@gmail.com>
* cluster: make sure we don't miss the first pushPull
During the join, memberlist initiates a pushPull to get initial data.
Unfortunately, at this point the nflog and silence listener have not
been registered yet, so the first data arrives only after one pushPull
cycle (1min by default !).
Signed-off-by: Corentin Chary <c.chary@criteo.com>
* Gossip large messages via SendReliable
For messages beyond half of the maximum gossip
packet size, send the message to all peer nodes
via TCP.
The choice of "larger than half the max gossip
size" is relatively arbitrary. From brief testing,
the overhead from memberlist on a packet seemed to
only use ~3 of the available 1400 bytes, and most
gossip messages seem to be <<500 bytes.
* Add tests for oversized/normal message gossiping
* Make oversize metric names consistent
* Remove errant printf in test
* Correctly increment WaitGroup
* Add comment for OversizedMessage func
* Add metric for oversized messages dropped
Code was added to drop oversized messages if the
buffered channel they are sent on is full. This
is a good thing to surface as a metric.
* Add counter for total oversized messages sent
* Change full queue log level to debug
Was previously a warning, which isn't necessary
now that there is a metric tracking it.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
When setting initially failing peers, if we don't
have a value for the advertise address, use the
bindAddr.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Correctly add Node to initially failed peer
Reconnect attempts to failed peers were panicking
because peer.Address() would attempt to access the
nil Node struct member.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Correctly remove old peers
Again, since we aren't assigning a name (this is
generated) we rely on the node's Address for
removing the initially joining (and potentially
later re-joining) peers
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Test the peerJoin removes initial peers
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Don't add self to failing peers list
The initially failing peers list shouldn't include
the bindAddr for the alertmanager itself, as this
connection is never made, and consequently only
removed from the failedPeers list after the failed
peer timeout.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Filter initialFailed with advertise addr
This may differ from bindAddr, and is the value we
want to not attempt to connect to.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Peers further propagate newly received nflogs
If a peer receives an nflog that it hasn't seen
before, queue the message and propagate it further
to other peers. This should ensure that all
peers within a cluster receive all gossip
messages.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Set Retransmit value based on number of members
For alertmanagers that are brought up with a list
of peers, set the number of message retransmits to
be half of that number. If there are no peers on
start, or there are few, continue to use the
default value of 3.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* [nflog] Move retransmit calculation
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* [silence] further gossip silence messages
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Set GossipNodes to equal RetransmitMulti
During a gossip, we send messages to at most
GossipNodes nodes. If possible, we only a message
to reach all nodes as soon as possible.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Fix rebase
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
memberlist doesn't advertise a valid IP address when the bind address is
empty (":8001") or the unspecified IPv6 address ("[::]:8001).
Signed-off-by: Simon Pasquier <spasquie@redhat.com>
* initial impl
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Add reconnectTimeout
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Fix locking
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Remove unused PeerStatuses
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Add metrics
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Actually use peerJoinCounter
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Cleanup peers map on peer timeout
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Add reconnect test
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* test removing failed peers
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Use peer address as map key
If a peer is restarted, it will rejoin with the
same IP but different ULID. So the node will
rejoin the cluster, but its peers will never
remove it from their internal list of failed nodes
because its ULID has changed.
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Add failed peers from creation
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Remove warnIfAlone()
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Update metric names
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
* Address comments
Signed-off-by: stuart nelson <stuartnelson3@gmail.com>
The cluster configuration uses DefaultLANConfig which seems
to be quite sensitive to WAN conditions. Allowing the tuning of these 3
parameters (TCP Timeout, Probe Interval and Probe Timeout) makes
clustering more robust across WAN connections.
Signed-off-by: Rhys Meaclem <rhysmeaclem@gmail.com>
* cluster: add alertmanager_cluster_messages_queued metric
* cluster: add metrics for sent messages
This change adds 2 new metrics:
- alertmanager_cluster_messages_sent_total
- alertmanager_cluster_messages_sent_size_total
* Fix marshaling for entries being broadcast
Individual notifications logs and silences being broadcast to the other
peers need to be encoded using the same length-delimited format as when
doing full-state synchronization.
* main: fix argument order for cluster.Join()
cluster.Join() was called with the push/pull and gossip interval
parameters being swapped one for another.
* Wait for the gossip to settle before sending notifications
See #1209 for details.
As an heuristic for mesh readyness, try to see if
the mesh looks stable (the number of peers isn't changing too much).
This implementation always mark the altermanager as ready after a maximum of 60s.
This adds one new flags to control this behavior:
```
--cluster.settle-timeout=60s mesh settling timeout. Do not wait more than this duration on startup.
```
It also adds `/-/ready` which always return 200 (in order to make it clear
that we are ready as soon as we can receive requests).
The mesh status is exposed in `/api/v1/status` and visible on `/#/status`.
* cluster: fix typos and base interval on gossipInterval
* cluster: gather alertmanager_peer_position all the time
This change moves the gathering of the alertmanager_peer_position metric
outside of the clusterWait() function so that the metric is computed
accurately even when no alerting group fires.
* cluster: add alertmanager_cluster_health_score metric
This metric is retrieved from the memberlist library.
gotjosh
Simon Pasquier
Pger-Y
Kien Nguyen-Tuan
Simon Pasquier
xuyasong
Povilas Versockas
Max Inden
Corentin Chary
stuart nelson
rhysm
Fabian Reinartz