Fix local docker invalid container name created by trimming (#958)

Co-authored-by: Yinghong Liu <[email protected]>

Author: YinghongLiu

torchx/schedulers/docker_scheduler.py

@@ -9,6 +9,7 @@
 import fnmatch
 import logging
 import os.path
+import re
 import tempfile
 from dataclasses import dataclass
 from datetime import datetime
@@ -265,9 +266,12 @@ def _submit_dryrun(self, app: AppDef, cfg: DockerOpts) -> AppDryRunInfo[DockerJo
                     rank0_env="TORCHX_RANK0_HOST",
                 )
                 replica_role = values.apply(role)
-                # trim app_id and role name in case name is longer than 64 letters. Assume replica_id is less than 10_000.
-                name = f"{app_id[-30:]}-{role.name[:30]}-{replica_id}"
-
+                # trim app_id and role name in case name is longer than 64 letters. Assume replica_id is less than 10_000. Remove invalid prefixes (https://github.com/moby/moby/blob/master/daemon/names/names.go#L6).
+                name = re.sub(
+                    r"^[^a-zA-Z0-9]+",
+                    "",
+                    f"{app_id[-30:]}-{role.name[:30]}-{replica_id}",
+                )
                 env = default_env.copy()
                 if replica_role.env:
                     env.update(replica_role.env)

torchx/schedulers/test/docker_scheduler_test.py

@@ -221,13 +221,15 @@ def test_long_hostname(self) -> None:
         for role in app.roles:
             role.name = "ethology_explore_magic_calliope_divisive_whirl_dealt_lotus_oncology_facet_deerskin_blum_elective_spill_trammel_trainer"
         with patch("torchx.schedulers.docker_scheduler.make_unique") as make_unique_ctx:
-            make_unique_ctx.return_value = "ethology_explore_magic_calliope_divisive_whirl_dealt_lotus_oncology_facet_deerskin_blum_elective_spill_trammel_12345"
+            make_unique_ctx.return_value = "ethology_explore_magic_calliope_divisive_whirl_dealt_lotus_oncology_facet_deerskin__.-_elective_spill_trammel_1234"
             info = self.scheduler.submit_dryrun(app, DockerOpts())
         for container in info.request.containers:
             assert "name" in container.kwargs
             name = container.kwargs["name"]
             assert isinstance(name, str)
             assert len(name) < 65
+            # Assert match container name rules https://github.com/moby/moby/blob/master/daemon/names/names.go#L6
+            self.assertRegex(name, r"[a-zA-Z0-9][a-zA-Z0-9_.-]")
 
 
 if has_docker():