Support for scaling but non-rolling deploys

docs/deploying.md

@@ -115,13 +115,19 @@ before starting the new one, resulting in a downtime of a few seconds
 This way, there is no risk of multiple containers writing to
 the same data at the same time.
 
-If `scaling` is given and larger than zero (so also when 1), a
+If `scale` is given and larger than zero (so also when 1), a
 zero-downtime deployment is possible, because the new containers will be
 started and given time to start up before the old containers are
 stopped. Note that in this case MyPaas assumes that the container is ready
 within 5s. You probably also want to specify `healthcheck` so that
 Traefik will not use a container before it is ready.
 
+If `scale` contains the word 'safe', the deployment is non-rolling. All old
+containers are stopped before the new ones are started, avoiding a client
+from obtaining a mix of old and new assets (especially important for
+apps that use a caching service worker). For symmetry, the word 'roll' can be
+used for rolling deploys.
+
 ### mypaas.healthcheck
 
 A value consisting of three values, e.g. "/status 10s 2s", representing

mypaas/server/_deploy.py

@@ -62,6 +62,7 @@ def get_deploy_generator(deploy_dir):
     port = 80
     portmaps = []
     scale = None
+    scale_option = "roll"
     urls = []
     volumes = []
     envvars = {}
@@ -98,6 +99,10 @@ def get_deploy_generator(deploy_dir):
                 elif key == "mypaas.publish":
                     portmaps.append(val)
                 elif key == "mypaas.scale":
+                    for opt in ("safe", "roll"):
+                        if opt in val:
+                            scale_option = opt
+                            val = val.replace(opt, "").strip()
                     scale = int(val)
                 elif key == "mypaas.healthcheck":
                     parts = val.split()
@@ -243,7 +248,14 @@ def label(x):
 
     # Deploy!
     if scale and scale > 0:
-        return _deploy_scale(container_infos, deploy_dir, service_name, cmd, scale)
+        if scale_option == "roll":
+            return _deploy_scale_roll(
+                container_infos, deploy_dir, service_name, cmd, scale
+            )
+        else:
+            return _deploy_scale_safe(
+                container_infos, deploy_dir, service_name, cmd, scale
+            )
     else:
         return _deploy_no_scale(container_infos, deploy_dir, service_name, cmd)
 
@@ -301,7 +313,7 @@ def _deploy_no_scale(container_infos, deploy_dir, service_name, prepared_cmd):
     yield f"done deploying {service_name}"
 
 
-def _deploy_scale(container_infos, deploy_dir, service_name, prepared_cmd, scale):
+def _deploy_scale_safe(container_infos, deploy_dir, service_name, prepared_cmd, scale):
     image_name = clean_name(service_name, ".-:/")
     base_container_name = clean_name(image_name, ".-")
 
@@ -315,6 +327,64 @@ def _deploy_scale(container_infos, deploy_dir, service_name, prepared_cmd, scale
     old_ids = get_id_name_for_this_service(container_infos)
     unique = str(int(time.time()))
 
+    yield f"renaming {len(old_ids)} current containers"
+    for i, id in enumerate(old_ids.keys()):
+        try:
+            dockercall("rename", id, base_container_name + f".old.{unique}.{i+1}")
+        except Exception:
+            yield "Rename failed. Probably a crashed container -> removing!"
+            dockercall("rm", "-f", id, fail_ok=True)
+
+    for id, name in old_ids.items():
+        yield f"stopping container (was {name})"
+        dockercall("stop", id, fail_ok=True)
+
+    # Keep track of started containers, in case we must shut them down
+    new_pool = []
+
+    try:
+        for i in range(scale):
+            new_name = f"{base_container_name}.{i+1}"
+            yield f"starting new container {new_name}"
+            new_pool.append(new_name)
+            cmd = prepared_cmd.copy()
+            cmd.append(f"--env=MYPAAS_CONTAINER={new_name}")
+            cmd.extend([f"--name={new_name}", image_name])
+            dockercall(*cmd)
+    except Exception:
+        yield "fail -> recovering"
+        for name in new_pool:
+            dockercall("stop", name, fail_ok=True)
+            dockercall("rm", name, fail_ok=True)
+        for id, name in old_ids.items():
+            dockercall("rename", id, name, fail_ok=True)
+            dockercall("start", id, fail_ok=True)
+        raise
+    else:
+        yield f"removing {len(old_ids)} old containers"
+        for id in old_ids.keys():
+            dockercall("rm", id, fail_ok=True)
+
+    yield "pruning"
+    dockercall("container", "prune", "--force")
+    dockercall("image", "prune", "--force")
+    yield f"done deploying {service_name}"
+
+
+def _deploy_scale_roll(container_infos, deploy_dir, service_name, prepared_cmd, scale):
+    image_name = clean_name(service_name, ".-:/")
+    base_container_name = clean_name(image_name, ".-")
+
+    yield ""
+    yield f"rolling deploy of {service_name} to containers {base_container_name}.1..{scale}"
+    time.sleep(1)
+
+    yield "building image"
+    dockercall("build", "-t", image_name, deploy_dir)
+
+    old_ids = get_id_name_for_this_service(container_infos)
+    unique = str(int(time.time()))
+
     yield f"renaming {len(old_ids)} current containers (and wait 2s)"
     for i, id in enumerate(old_ids.keys()):
         try:

mypaas/server/_traefik.py

@@ -139,6 +139,7 @@ def init_router():
   middlewares = ["auth"]
   [http.routers.api.tls]
     certresolver = "default"
+    options= "intermediate"
 
 # The routing for mypaas daemon
 [http.routers.mypaas-daemon-router]
@@ -147,6 +148,7 @@ def init_router():
   service = "mypaas-daemon"
   [http.routers.mypaas-daemon-router.tls]
     certresolver = "default"
+    options= "intermediate"
 [http.services.mypaas-daemon.loadBalancer]
     [[http.services.mypaas-daemon.loadBalancer.servers]]
       url = "http://127.0.0.1:88"