Gateway state sync mechanism (#4135)

Introduce a way to reconcile gateway replica state
based on the current list of services. This allows
newly started or newly recovered gateway replicas
to learn about existing services. It also improves
fault tolerance, since an unavailable gateway
replica no longer blocks the service lifecycle -
services and service replicas can still be
registered on the remaining healthy gateway
replicas.

Behavioral details:
- Run submission no longer waits for service
  registration before the run starts provisioning.
  This has both advantages (provisioning and
  service registration can happen in parallel) and
  disadvantages (compute may be provisioned for a
  service that will be terminated a moment later
  due to a registration error). This behavior may
  be revised in the future.
- A service or a service replica is terminated
  with the `gateway_error` termination reason if
  it failed to be registered on all gateway
  replicas.
- If registration succeeded at least on one
  gateway replica, registration on the remaining
  gateway replicas will be retried indefinitely.
- Rolling deployments only proceed with
  terminating a service replica once the
  replacement replica is successfully registered
  on all gateway replicas.
- Registration and unregistration tracking,
  including per-gateway-replica registration error
  messages, is available in events.

Author: jvstme

mkdocs/docs/concepts/gateways.md

@@ -174,7 +174,6 @@ $ dstack gateway list
 
     - Changing the number of replicas or redeploying replicas is not supported.
     - HTTPS is only supported for AWS gateways with the `acm` [certificate type](#certificate). For other gateways, use an external load balancer for TLS termination.
-    - An unavailable gateway replica prevents any new services or service replicas from being added.
     - All replicas are bound to the same backend and region.
     - At most 3 replicas are allowed per gateway.
 

src/dstack/_internal/core/models/runs.py

@@ -94,6 +94,7 @@ class RunTerminationReason(str, Enum):
     STOPPED_BY_USER = "stopped_by_user"
     ABORTED_BY_USER = "aborted_by_user"
     SERVER_ERROR = "server_error"
+    GATEWAY_ERROR = "gateway_error"
 
     def to_job_termination_reason(self) -> "JobTerminationReason":
         """
@@ -107,6 +108,7 @@ def to_job_termination_reason(self) -> "JobTerminationReason":
             self.STOPPED_BY_USER: JobTerminationReason.TERMINATED_BY_USER,
             self.ABORTED_BY_USER: JobTerminationReason.ABORTED_BY_USER,
             self.SERVER_ERROR: JobTerminationReason.TERMINATED_BY_SERVER,
+            self.GATEWAY_ERROR: JobTerminationReason.TERMINATED_BY_SERVER,
         }
         return mapping[self]
 
@@ -118,6 +120,7 @@ def to_status(self) -> "RunStatus":
             self.STOPPED_BY_USER: RunStatus.TERMINATED,
             self.ABORTED_BY_USER: RunStatus.TERMINATED,
             self.SERVER_ERROR: RunStatus.FAILED,
+            self.GATEWAY_ERROR: RunStatus.FAILED,
         }
         return mapping[self]
 
@@ -126,6 +129,8 @@ def to_error(self) -> Optional[str]:
             return "retry limit exceeded"
         elif self == RunTerminationReason.SERVER_ERROR:
             return "server error"
+        elif self == RunTerminationReason.GATEWAY_ERROR:
+            return "gateway error"
         else:
             return None