What happened?
What happened?
Pillar (and any other) requests made via salt-call on a minion take ~60 seconds
(the full REQUEST_TIMEOUT) to complete on 3006.27, when the same command
completes in a few seconds on 3006.26. The affected code path is unchanged
in 3006.26, so this is a regression introduced in 3006.27.
[06:31:40] alma98-p3g ~ # time salt-call pillar.data dontcarejustatest
local:
----------
dontcarejustatest:
real 1m3.404s
user 0m1.045s
sys 0m0.137s
Trace logging on the minion shows the request hangs for ~60s waiting on a
reply from the master before finally timing out and retrying successfully:
[DEBUG ] Initializing new AsyncAuth for (...)
[TRACE ] ext_pillar_extra_data = {}
[DEBUG ] salt.crypt.get_rsa_key: Loading private key
<-- ~60s stall -->
[TRACE ] Failed to send msg SaltReqTimeoutError('Message timed out')
[TRACE ] Request timed out while waiting for a response. reconnecting.
Root cause
3006.27 includes commit 36e914a72af330e8e0c09bfea5f43893f5e972b1
("Use stable ZMQ identity for minion + syndic daemon ret-port REQ
sockets"), which sets a stable ZMQ IDENTITY on the AsyncReqMessageClient
REQ socket for long-lived minion/syndic daemons:
elif _role in ("minion", "syndic") and _minion_id:
identity = "salt-req/{role}/{minion_id}/{slot}".format(
role=_role,
minion_id=_minion_id,
slot=next(_REQ_IDENTITY_SLOT),
)
self.socket.setsockopt(zmq.IDENTITY, identity.encode("utf-8"))
_REQ_IDENTITY_SLOT is a module-level itertools.count() that starts at 0
per process. A running salt-minion daemon and a one-off salt-call
invocation both set opts["__role"] = "minion" and share the same
minion_id, so both processes independently compute the identity
salt-req/minion/<minion_id>/0 for their first REQ socket.
Because the master's ROUTER socket has ROUTER_HANDOVER=1 set (see
salt/transport/zeromq.py, MWorkerQueue), when the second process
(salt-call) connects with the same identity as the first (salt-minion),
libzmq hands the peer slot over to the newcomer and silently drops any
reply still in flight to the original owner. This causes whichever process
was waiting on that reply to stall for the full REQUEST_TIMEOUT (60s)
before retrying.
This directly contradicts the stated goal of the commit itself, which
explicitly calls out that giving two sockets the same identity causes
exactly this failure mode ("Giving them all the same stable identity would
cause ROUTER_HANDOVER on the upstream ROUTER to silently drop any reply
still in flight..."), but the fix only accounted for multiple sockets
within one process, not multiple processes (salt-minion and
salt-call) sharing the same minion_id and hence the same identity
namespace.
Suggested fix
Include something process-unique (e.g. os.getpid()) in the identity
string for the minion/syndic branch, so that salt-call,
salt-minion, and any other short-lived process using the minion config
never collide with the long-running daemon's identity slots.
Type of salt install
Official rpm
Major version
3006.x
What supported OS are you seeing the problem on? Can select multiple. (If bug appears on an unsupported OS, please open a GitHub Discussion instead)
almalinux-9
salt --versions-report output
salt-minion 3006.27 (Sulfur)
What happened?
What happened?
Pillar (and any other) requests made via
salt-callon a minion take ~60 seconds(the full
REQUEST_TIMEOUT) to complete on 3006.27, when the same commandcompletes in a few seconds on 3006.26. The affected code path is unchanged
in 3006.26, so this is a regression introduced in 3006.27.
Trace logging on the minion shows the request hangs for ~60s waiting on a
reply from the master before finally timing out and retrying successfully:
Root cause
3006.27 includes commit
36e914a72af330e8e0c09bfea5f43893f5e972b1("Use stable ZMQ identity for minion + syndic daemon ret-port REQ
sockets"), which sets a stable ZMQ
IDENTITYon theAsyncReqMessageClientREQ socket for long-lived
minion/syndicdaemons:_REQ_IDENTITY_SLOTis a module-levelitertools.count()that starts at 0per process. A running
salt-miniondaemon and a one-offsalt-callinvocation both set
opts["__role"] = "minion"and share the sameminion_id, so both processes independently compute the identitysalt-req/minion/<minion_id>/0for their first REQ socket.Because the master's ROUTER socket has
ROUTER_HANDOVER=1set (seesalt/transport/zeromq.py,MWorkerQueue), when the second process(
salt-call) connects with the same identity as the first (salt-minion),libzmq hands the peer slot over to the newcomer and silently drops any
reply still in flight to the original owner. This causes whichever process
was waiting on that reply to stall for the full
REQUEST_TIMEOUT(60s)before retrying.
This directly contradicts the stated goal of the commit itself, which
explicitly calls out that giving two sockets the same identity causes
exactly this failure mode ("Giving them all the same stable identity would
cause ROUTER_HANDOVER on the upstream ROUTER to silently drop any reply
still in flight..."), but the fix only accounted for multiple sockets
within one process, not multiple processes (
salt-minionandsalt-call) sharing the sameminion_idand hence the same identitynamespace.
Suggested fix
Include something process-unique (e.g.
os.getpid()) in the identitystring for the
minion/syndicbranch, so thatsalt-call,salt-minion, and any other short-lived process using the minion confignever collide with the long-running daemon's identity slots.
Type of salt install
Official rpm
Major version
3006.x
What supported OS are you seeing the problem on? Can select multiple. (If bug appears on an unsupported OS, please open a GitHub Discussion instead)
almalinux-9
salt --versions-report output