Skip to content

Kolla images: start path and one-shot containers

This page answers open question Q14 of the Open questions: what the kolla images do when a container starts, and what the one-shot containers that Kolla-Ansible starts with KOLLA_BOOTSTRAP, KOLLA_UPGRADE, KOLLA_OSM and similar variables actually run. It was read from the kolla image repository (openstack/kolla, tag ussuri-eol, commit 9107c5f) and matched against Kolla-Ansible ussuri-eol (8a94fe9). The images were read, not built or run.

Note

Key answers

  • What every container start does: kolla_start runs kolla_set_configs, kolla_copy_cacerts and the image's kolla_extend_start, then execs the service command (§2.2).
  • kolla_set_configs --check compares only the config_files entries of config.json (content, mode, owner) CODE (§3.1). It never compares command, permissions or removed entries (§3.2). Exit codes 0, 1, 2 and 137, and what kolla_docker does with each: §3.3. Run by hand, it needs docker exec -u root (§3.4).
  • What a COPY_ALWAYS restart applies: changed or added files, mode and owner changes and a changed command CODE; removals do not apply (§3.5).
  • One-shot containers: what KOLLA_BOOTSTRAP, KOLLA_UPGRADE, KOLLA_OSM and the other variables run in each image, and which Kolla-Ansible task sets them (§4).
  • Per-image start behaviour: the table in §5. The nova-libvirt start path is passive towards running guests (§5.1); with enable_neutron_agent_ha, the L3 agent wrapper runs neutron-netns-cleanup --force --agent-type l3 on every container start (§5.3).
  • Day-2 consequences: 19 items (§6). What is not available from the sources: §7.

Citation convention: kolla image paths carry a kolla prefix (kolla docker/base/start.sh:8). Kolla-Ansible paths have no prefix and are relative to its ansible/ directory (roles/nova/tasks/bootstrap_service.yml:13, library/kolla_docker.py:543-578, nova.yml:45-61), except tools/.... Ansible source paths carry an ansible 2.9.27 prefix. Evidence labels are those of Conventions and evidence labels: CODE (a kolla or Ansible path counts as CODE) and NOT AVAILABLE (the sources do not establish it).


1. Scope and method

  • Read in full CODE: kolla docker/base (Dockerfile.j2, start.sh, set_configs.py, copy_cacerts.sh, httpd_setup.sh, sudoers, healthcheck_*), and the Dockerfile.j2 and start scripts of the images in §5: Keystone, Glance, Placement, Cinder, Heat, Horizon, Barbican, Octavia, Manila, Swift, Nova, Neutron, Open vSwitch, OVN, MariaDB, RabbitMQ, memcached, HAProxy, keepalived (with check_alive.sh), fluentd, cron, kolla-toolbox, chrony, iscsid and tgtd. Of the other images only the ironic-api bootstrap branch (§4) and the bifrost-deploy ENTRYPOINT/CMD (§2.1) were read, and every Dockerfile was searched for HEALTHCHECK (§2.7) and for ENTRYPOINT/CMD overrides. The rest were not.
  • Matched to Kolla-Ansible. Each one-shot variable was traced to the task that sets it (§4) CODE.
  • kolla_set_configs --check was read, not run: exit codes 0, 1 and 2 of §3.3 are read from kolla docker/base/set_configs.py; 137 and the exec client error are kolla_docker outcomes (library/kolla_docker.py:543-578).
  • Not covered: behaviour against real services and data, signal delivery, timing. These are NOT AVAILABLE and are listed in §7.

2. Start path common to every image

2.1 Image wiring

Every kolla image inherits from kolla docker/base/Dockerfile.j2 CODE:

  • ENTRYPOINT ["dumb-init", "--single-child", "--"] when use_dumb_init is on (:433-435; default True, kolla kolla/common/config.py:170), and CMD ["kolla_start"] (:455). The MariaDB image drops --single-child (§2.6). bifrost-deploy resets both (ENTRYPOINT [], CMD [ "/sbin/init" ], kolla docker/bifrost/bifrost-deploy/Dockerfile.j2:29-30), so bifrost_deploy never runs kolla_start or kolla_set_configs.
  • The base scripts are installed as kolla_start (start.sh), kolla_set_configs (set_configs.py), kolla_copy_cacerts (copy_cacerts.sh) and kolla_httpd_setup (httpd_setup.sh); sudoers becomes /etc/sudoers, mode 0440 (:427-431,448).
  • /usr/local/bin/kolla_extend_start is created empty by touch (:446). An image gets start logic only by overwriting it.
  • /var/log/kolla gets group kolla and mode 2775 (:449-451).
  • Service users are added to group kolla (kolla docker/macros.j2:50). The base sudoers lets members of that group run two commands as root without a password, keeping their environment: kolla_set_configs and kolla_copy_cacerts (kolla docker/base/sudoers:7,16,19). It also includes /etc/sudoers.d (:21), where many images add entries of their own, e.g. cinder-volume's chown -R and chmod (kolla docker/cinder/cinder-volume/cinder_sudoers:2-5) and mariadb's kolla_security_reset (kolla docker/mariadb/mariadb/mariadb_sudoers:1). The kolla_set_configs entry restricts no arguments, so sudo kolla_set_configs --check is allowed.
  • Kolla-Ansible passes KOLLA_CONFIG_STRATEGY to every container through docker_common_options (group_vars/all.yml:43,162-163; default COPY_ALWAYS).

2.2 kolla_start step by step

flowchart TB
  D["dumb-init --single-child (PID 1)"] --> KS["kolla_start (errexit, xtrace)"]
  KS --> SC["sudo -E kolla_set_configs"]
  SC --> ST{"KOLLA_CONFIG_STRATEGY"}
  ST -- "COPY_ALWAYS" --> CA["copy + permissions + /run_command (every start)"]
  ST -- "COPY_ONCE, first start" --> CO["copy + permissions + /run_command, create /configured"]
  ST -- "COPY_ONCE, /configured exists" --> SK["copy nothing, exit 0"]
  ST -- "invalid strategy or copy error" --> F1["exit 1 or 2: start aborts"]
  CA --> RC["CMD = cat /run_command"]
  CO --> RC
  SK --> RC
  RC --> CC["sudo kolla_copy_cacerts"]
  CC --> SE{"KOLLA_SKIP_EXTEND_START set?"}
  SE -- "yes" --> EX
  SE -- "no" --> XS["source kolla_extend_start (image USER)"]
  XS --> OS{"one-shot variable set?"}
  OS -- "yes" --> Z["run DB or setup commands, exit 0"]
  OS -- "no" --> EX["exec CMD ARGS"]
  1. set -o errexit; set -o xtrace (kolla docker/base/start.sh:2-3). Any failing command ends the container, including commands in the sourced kolla_extend_start. xtrace is on in that script, and in the kolla_<service>_extend_start it sources (§2.5), where some one-shots put a secret on a command line: heat's --password ${HEAT_DOMAIN_ADMIN_PASSWORD} (kolla docker/heat/heat-api/extend_start.sh:12) and the MariaDB root password (kolla docker/mariadb/mariadb/extend_start.sh:23-25) CODE. NOT AVAILABLE (bash semantics, not in these sources) Whether the trace on stderr shows those commands with their variables expanded, and so writes the secret there.

    kolla_docker returns the container's stderr in the task result (library/kolla_docker.py:832-836,840-845,1070). Ansible's default callback prints a failed result, stderr included, and an ok or changed one only with -v (ansible 2.9.27 lib/ansible/plugins/callback/default.py:94-114,152-153, lib/ansible/plugins/callback/__init__.py:101-123); a task with no_log reaches the callback with its result replaced by a censored message (ansible 2.9.27 lib/ansible/executor/task_queue_manager.py:373, lib/ansible/executor/task_result.py:131-139). bootstrap_heat and bootstrap_mariadb set no no_log (roles/heat/tasks/bootstrap_service.yml:2-29, roles/mariadb/tasks/bootstrap_cluster.yml:2-23), upgrade_mariadb does (roles/mariadb/tasks/upgrade.yml:26) CODE.

  2. sudo -E kolla_set_configs (:8) runs as root whatever the image USER, copies the configuration (§2.3) and writes config.json's command to /run_command.

  3. CMD=$(cat /run_command) (:9).
  4. sudo kolla_copy_cacerts (:13, §2.4), on every start whatever the strategy.
  5. Unless KOLLA_SKIP_EXTEND_START is set, . kolla_extend_start (:15-18). The script is sourced: it runs in the same shell, as the image USER, with errexit on. It can set ARGS or CMD, or exit. An exit 0 ends the container before the service starts, which is how every one-shot branch in §4 works. The test is ${!KOLLA_SKIP_EXTEND_START[@]}, the idiom that kolla's own comments describe as catching "all cases of the ... variable being set, including empty" (kolla docker/keystone/keystone/extend_start.sh:40-42), so any value, even empty or false, skips the script and every one-shot branch with it CODE. Kolla-Ansible never sets this variable (grep of ansible/).
  6. exec ${CMD} ${ARGS}, unquoted (:20-21). The service replaces the shell and becomes dumb-init's only child.

2.3 kolla_set_configs: copy and permissions

Without arguments, kolla_set_configs loads config.json (the KOLLA_CONFIG variable as a JSON string if set, otherwise KOLLA_CONFIG_FILE, default /var/lib/kolla/config_files/config.json), validates it (command is required; each config_files entry needs source and dest, plus owner and perm unless preserve_properties) and applies the strategy (kolla docker/base/set_configs.py:238-253,272-307,388-404) CODE:

  • COPY_ALWAYS (:391-393): copy and permissions on every start.
  • COPY_ONCE (:394-402): on the first start, copy, permissions and mknod /configured. On any later start of the same container ImmutableConfig is raised with exit code 0: nothing is copied, permissions are not re-applied and /run_command keeps its first value. kolla_start continues normally.
  • Unset or any other value: InvalidConfig, exit 1 (:403-404).

Copy, per config_files entry (ConfigFile.copy, :131-158):

  • The source is a glob. A non-optional entry that matches nothing raises MissingRequiredSource (exit 1). An optional one is skipped silently (:133-139).
  • A dest ending in / means dest/<basename of source> (:145-146).
  • Unless merge is set, the destination is deleted first, a whole tree for a directory, then parent directories are created (:147-149). No Kolla-Ansible config.json template uses merge (grep of roles/*/templates/*.json.j2).
  • Directory sources are copied recursively. Every directory and file gets the entry's owner and perm (so directories get the file mode), or the source's mode and owner with preserve_properties (:85-99,115-129).
  • An OSError during the copy is logged and re-raised: exit 2 (:150-158,430-432).
  • Finally command is written to /run_command, mode 0644 (:319-329).

Permissions, per permissions entry (handle_permissions, :342-385):

  • path is a glob. A path that matches nothing is skipped without a warning (:378).
  • owner is user[:group], the group defaulting to the user (:332-339). An unknown user or group raises KeyError: exit 2 (:349-351,430-432).
  • chown always, chmod only if perm is given. A failing chown or chmod is only logged (:359-376).
  • With recurse, the same is applied to every directory and file below, one unconditional os.chown per entry (:380-385). Several manifests do this on data volumes (§6 item 10).

A failure here (missing required source, bad JSON, unknown user) aborts kolla_start through errexit. The default restart policy is unless-stopped (group_vars/all.yml:134) CODE. NOT AVAILABLE Whether the container then restarts in a loop until the input is fixed.

2.4 kolla_copy_cacerts and kolla_httpd_setup

  • kolla_copy_cacerts deletes every kolla-customca-* file from the system trust store, copies each file of /var/lib/kolla/config_files/ca-certificates/ back under that prefix, and runs update-ca-certificates (Debian/Ubuntu) or update-ca-trust (CentOS/RHEL) (kolla docker/base/copy_cacerts.sh:7-27) CODE. The update command runs only when that directory exists and is not empty (:10-11). NOT AVAILABLE Whether, once the last custom CA is removed, the generated trust bundle in the container layer keeps it until the container is recreated.
  • kolla_httpd_setup acts only when the shell runs as root (whoami, kolla docker/base/httpd_setup.sh:7). It empties /var/run/apache2/ (Debian/Ubuntu) or removes /var/run/httpd/*, /run/httpd/* and /tmp/httpd*, so a restarted httpd does not trip on stale run files. On CentOS it also creates localhost.crt if missing (:10-26). In images that run as a service user (nova-api, barbican-api, octavia-api, manila-api) it does nothing.

2.5 The extend_start chain

Most service families layer two scripts CODE. The service base image installs its extend_start.sh as kolla_extend_start, creates an empty /usr/local/bin/kolla_<service>_extend_start, and its last line sources that file (e.g. kolla docker/cinder/cinder-base/Dockerfile.j2:73-79, docker/cinder/cinder-base/extend_start.sh:10). The API image overwrites kolla_<service>_extend_start with the one-shot branches. On those images kolla_extend_start prepares the log directory (neutron-base also selects the iptables backend, §5), and the DB commands are in kolla_<service>_extend_start. Keystone, keystone-fernet, keystone-ssh, horizon, mariadb, rabbitmq, swift-rsyncd, keepalived, chrony, iscsid, fluentd, ovn-base, nova-libvirt and nova-ssh put everything in kolla_extend_start instead.

2.6 Signals and PID 1

  • PID 1 is dumb-init --single-child, and kolla_start execs the service, so the service is dumb-init's direct child (kolla docker/base/Dockerfile.j2:434, docker/base/start.sh:21) CODE. NOT AVAILABLE (dumb-init semantics, not in these sources) Whether in this mode dumb-init forwards docker stop's SIGTERM only to that child. A comment in the mariadb image describes the other mode: without --single-child, "the TERM signal is sent to all subprocesses" (kolla docker/mariadb/mariadb/Dockerfile.j2:67-71).
  • MariaDB overrides the entrypoint to dumb-init -- without --single-child (kolla docker/mariadb/mariadb/Dockerfile.j2:65-76) CODE. Its comment gives the reason: mysqld_safe ignores SIGTERM, and without --single-child TERM also reaches mysqld CODE (for the comment; NOT AVAILABLE: dumb-init and mysqld_safe semantics).
  • Some Kolla-Ansible commands are shell wrappers that do not exec the service, so dumb-init's child is bash: neutron-l3-agent-wrapper.sh (roles/neutron/templates/neutron-l3-agent-wrapper.sh.j2:14-18), start-ovs and start-ovsdb-server (roles/openvswitch/templates/start-ovs.j2:2, start-ovsdb-server.j2:11) and haproxy_run.sh (roles/haproxy/templates/haproxy_run.sh.j2:7-9) CODE. NOT AVAILABLE Whether the service then gets any SIGTERM before the container stops. kolla_docker stops containers with the Docker client's stop and a 10 s timeout (library/kolla_docker.py:871-884,892-904, Repository architecture §6) CODE.

2.7 Healthcheck scripts

The base image installs healthcheck_curl (curl --fail, 10 s maximum), healthcheck_filemod, healthcheck_listen, healthcheck_port and healthcheck_socket when docker_healthchecks is on (kolla docker/base/Dockerfile.j2:437-444; default True, kolla kolla/common/config.py:253). No kolla Dockerfile has a HEALTHCHECK instruction, and nothing in Kolla-Ansible ussuri-eol configures a container healthcheck (grep of ansible/ for healthcheck hits only monasca and swift service settings) CODE. Ussuri containers therefore have no Docker health status. The mariadb image's own healthcheck_mariadb exits 0 whether the node is synced or not (kolla docker/mariadb/mariadb/healthcheck_mariadb:10-17).


3. kolla_set_configs --check: what it compares and its exit codes

kolla_docker's compare_container runs /usr/local/bin/kolla_set_configs --check inside the running container with docker exec as root, and its exit code decides config drift (library/kolla_docker.py:244,326-332,543-578; Repository architecture §4.3). --check loads and validates config.json like a normal start, then calls ConfigFile.check for each config_files entry (kolla docker/base/set_configs.py:407-426).

3.1 What is compared

For each entry whose source glob matches something (:213-235) CODE:

  • File source (_cmp_file, :160-191): destination missing on a non-optional entry (drift); byte-for-byte content; mode, as the last four octal digits of st_mode against the entry's perm string; owner user and group names.
  • Directory source (_cmp_dir, :193-211): every file below, as above. For subdirectories the relpath arguments are reversed (:197-198), so the mode checked is that of the destination's parent against the file perm. A directory source that contains a subdirectory therefore reports drift on every run.
  • preserve_properties entries carry no perm (:60-65), so the mode test always fails (:175) and they report drift on every run. Example: prometheus_server's extras/* entry whenever extras exist (roles/prometheus/templates/prometheus-server.json.j2:10-15).

3.2 What is never compared

  • config.json's command and permissions: --check iterates only config_files (:407-410). A manifest with "config_files": [], like memcached's, whose settings are all in its command, always returns 0 (roles/memcached/templates/memcached.json.j2:2-3).
  • Entries removed from config.json, and optional entries whose source no longer exists (:219-220), even while their old destination is still in the container.
  • Extra files in a destination directory, and the mode and owner of the destination directories themselves (for subdirectories _cmp_dir checks the mode of a directory above the destination instead, §3.1). A file destination's mode and owner are compared (§3.1).

3.3 Exit codes

Exit code Raised when What kolla_docker does Day-2 meaning
0 Every listed entry matches. Also a manifest with empty config_files, and any difference confined to command, permissions, removed entries or optional entries without a source no drift (kolla_docker.py:566-567) "No drift" covers the listed files only. A config.json-only change written by genconfig makes neither deploy-containers nor a later run restart the container (§6 item 1)
1 ConfigFileBadState: a listed file differs in content, mode or owner, or a non-optional destination is missing. Also MissingRequiredSource (a non-optional source matches nothing, set_configs.py:217-218) and InvalidConfig (bad JSON, no command, an entry without owner/perm; :238-253,282,295) drift: the handler restarts the container (kolla_docker.py:568-569) Usually real drift. For a missing source or an invalid config.json the restarted container then fails in kolla_set_configs at every start (errexit; whether it then restarts in a loop is NOT AVAILABLE, §2.3). preserve_properties entries and directory sources with subdirectories return 1 on every run
2 Any other exception (set_configs.py:430-432): an optional entry whose source exists but whose destination is not in the container yet (open() raises at :167); a config.json that cannot be opened (:291, e.g. run as a service user); a uid or gid with no passwd or group entry raises: the check task fails on that host (kolla_docker.py:575-578,1071-1073) Real drift reported as an error, e.g. a newly rendered nova-compute release file (roles/nova-cell/templates/nova-compute.json.j2:52-57). The run fails instead of restarting the container. Entries are checked in order and the first difference ends the check (set_configs.py:407-410,427-429), so if an entry listed before the optional one differs, the result is 1 and the restart also copies the new file
137 the exec was killed because the container exited drift (kolla_docker.py:570-574) the container is unstable; Kolla restarts it
none docker exec fails with a client error (container stopped or restarting) drift (kolla_docker.py:552-558) a stopped container is restarted by the next run

Read from main (kolla docker/base/set_configs.py:413-433) CODE: rc 0 for a match, for empty config_files with a changed command (:407-410), for an optional entry without a source (:219-220), and for a flat directory with an extra file in the destination (_cmp_dir walks the source only, :193-211); rc 1 for a content, mode or missing required destination difference (:162-178), a missing required source (:217-218), a preserve_properties entry (:175), a directory with a subdirectory (:197-198), a missing command (:241-242) and bad JSON (:294-296); rc 2 for a new optional file (:167) and a missing config.json (:291).

3.4 Running it by hand needs root

Kolla-Ansible writes the files in /etc/kolla/<service>/ with mode 0660 into directories with mode 0770, owned by config_owner_user root (group_vars/all.yml:20-21; e.g. roles/nova/tasks/config.yml:2-9, roles/keystone/tasks/config.yml:45-50), and mounts them read-only (e.g. roles/nova-cell/defaults/main.yml:284). In an image with a service USER (e.g. USER nova, kolla docker/nova/nova-compute/Dockerfile.j2:196) a plain docker exec <c> kolla_set_configs --check cannot open them and exits 2. kolla_docker execs as root (library/kolla_docker.py:545-549). By hand, use docker exec -u root as the swift reconfigure does (roles/swift/tasks/reconfigure.yml:47), or sudo kolla_set_configs --check inside the container (§2.1). The bifrost reconfigure omits -u root (roles/bifrost/tasks/reconfigure.yml:17), but its check cannot work anyway: it execs into bifrost-deploy, while the container is bifrost_deploy (roles/bifrost/tasks/start.yml:8), which runs /sbin/init and mounts no config.json (§2.1). No kolla-ansible command runs that file.

3.5 What a restart applies

  • A COPY_ALWAYS restart (stop and start of the same container, library/kolla_docker.py:892-904) re-copies the listed entries and rewrites /run_command. Changed or added files, mode and owner changes and a changed command apply CODE. Stale files inside a listed directory destination go too, because the destination is deleted first.
  • Removals do not apply. The destinations of entries dropped from config.json, and of optional sources that are gone, are never deleted (kolla docker/base/set_configs.py:135-148,310-315). They stay in the container's writable layer until the container is recreated; a destination on a volume survives recreates as well, e.g. nova-compute's SASL auth.conf (roles/nova-cell/templates/nova-compute.json.j2:58-64), left behind when libvirt_enable_sasl is turned off, is under /var/lib/nova, by default the nova_compute volume (roles/nova-cell/defaults/main.yml:293, group_vars/all.yml:967). --check does not flag them CODE. NOT AVAILABLE Whether Keystone still reads a removed policy file (roles/keystone/templates/keystone.json.j2:30-36) after the restart.
  • CA certificates are purged and re-added on every start (§2.4), from /etc/kolla/<svc>/ca-certificates/. A CA deleted from kolla_certificates_dir/ca/, or left there after kolla_copy_ca_into_containers is turned off, stays in that directory, because the copy task never deletes files (roles/service-cert-copy/tasks/main.yml:2-12; the copy action only walks the source tree, ansible 2.9.27 lib/ansible/plugins/action/copy.py:485-582) CODE, so it is re-added at every start until it is deleted there by hand. NOT AVAILABLE Whether, once the last custom CA is removed there, the generated trust bundle keeps it (§2.4).
  • COPY_ONCE: /configured lives in the container layer, so only a recreate re-applies config. kolla_docker always recreates under COPY_ONCE (library/kolla_docker.py:793-801).

4. One-shot containers: what the environment variables run

Kolla-Ansible starts one-shot containers with kolla_docker start_container, detach: False, restart_policy: no, a label such as BOOTSTRAP or UPGRADE, and a variable that selects a branch of the image's start script. The variables are sent with an empty value (library/kolla_docker.py:758-760) and every branch tests only whether they are set (${!VAR[@]}). The one-shot runs the whole of kolla_start first: kolla_set_configs with the config directory mounted into it, kolla_copy_cacerts, then the branch, which ends in exit 0 before the service command (§2.2) CODE. That directory is the service's own, except for Nova: its one-shots mount a dedicated nova-api-bootstrap or nova-cell-bootstrap directory whose command is "false" (roles/nova/defaults/main.yml:116-117, roles/nova-cell/defaults/main.yml:303-304).

Image KOLLA_BOOTSTRAP KOLLA_UPGRADE / other one-shot variables Set by Kolla-Ansible (task, action) Day-2 note
keystone keystone-manage db_sync, as user keystone (kolla docker/keystone/keystone/extend_start.sh:43-46) KOLLA_UPGRADE: db_sync --expand, then --migrate (:24-31). KOLLA_FINISH_UPGRADE: db_sync --contract (:36-39) bootstrap_keystone, deploy/reconfigure, run_once (roles/keystone/tasks/bootstrap_service.yml:32-49). Handlers init_upgrade_database and finish_upgrade_database, upgrade only (roles/keystone/handlers/main.yml:5-27,74-95) Schema sync on every deploy/reconfigure. EXTRA_KEYSTONE_MANAGE_ARGS (:21) is never set by Kolla-Ansible
glance-api glance-manage db_sync, then db_load_metadefs (kolla docker/glance/glance-api/extend_start.sh:5-9) KOLLA_UPGRADE with GLANCE_DB_EXPAND, GLANCE_DB_MIGRATE or GLANCE_DB_CONTRACT: glance-manage db expand, migrate or contract (:13-27) bootstrap_glance, deploy/reconfigure and legacy upgrade (roles/glance/tasks/bootstrap_service.yml:18-36, legacy_upgrade.yml:6). Rolling containers only with glance_enable_rolling_upgrade (default no, group_vars/all.yml:868; rolling_upgrade.yml:29-101) glance-manage db_load_metadefs runs on every deploy/reconfigure; whether, without --merge, it only adds namespaces not yet in the DB is NOT AVAILABLE
placement-api placement-manage db sync, then db online_data_migrations (kolla docker/placement/placement-api/extend_start.sh:5-9) KOLLA_UPGRADE: db sync only (:11-15), never used. KOLLA_OSM: db online_data_migrations (:17-21) bootstrap_placement, deploy/reconfigure and upgrade (roles/placement/tasks/bootstrap_service.yml:2-20, deploy.yml:11, reconfigure.yml:2, upgrade.yml:11). KOLLA_OSM after the restart, upgrade only (upgrade.yml:16-34) Online data migrations on every deploy/reconfigure, and twice on upgrade
cinder-api cinder-manage db sync (kolla docker/cinder/cinder-api/extend_start.sh:6-9) No KOLLA_UPGRADE. KOLLA_OSM: db online_data_migrations, with --max_count only if MAX_NUMBER is set (:11-18) bootstrap_cinder, deploy/reconfigure and upgrade (roles/cinder/tasks/bootstrap_service.yml:2-20, upgrade.yml:7). KOLLA_OSM without MAX_NUMBER, so unbounded, upgrade only (upgrade.yml:16-34) A tag change applied by deploy/reconfigure syncs the schema without the online migrations
heat-api heat-manage db_sync, then, only if heat_user_domain is absent, creates that domain, user heat_domain_admin (password HEAT_DOMAIN_ADMIN_PASSWORD) and its admin role (kolla docker/heat/heat-api/extend_start.sh:5-17) none; upgrade re-runs the bootstrap bootstrap_heat with admin OS_* credentials, deploy/reconfigure and upgrade (roles/heat/tasks/bootstrap_service.yml:10-21, upgrade.yml:6) Needs a reachable Keystone. A changed heat_domain_admin_password never reaches Keystone (§6 item 7)
horizon manage.py migrate --noinput, after the settings symlinks (kolla docker/horizon/extend_start.sh:24-49) none; FORCE_GENERATE is a start-time switch (§5.2) bootstrap_horizon only when horizon_backend_database (= enable_murano, group_vars/all.yml:982; roles/horizon/tasks/deploy.yml:10-11, upgrade.yml:10-11) Usually never runs
barbican-api barbican-db-manage upgrade (kolla docker/barbican/barbican-api/extend_start.sh:5-8) none bootstrap_barbican, deploy/reconfigure and upgrade (roles/barbican/tasks/bootstrap_service.yml:2-20, upgrade.yml:6) Schema upgrade on every deploy/reconfigure
octavia-api octavia-db-manage upgrade head (kolla docker/octavia/octavia-api/extend_start.sh:5-8) none bootstrap_octavia, deploy/reconfigure and upgrade (roles/octavia/tasks/bootstrap_service.yml:2-20, upgrade.yml:6) Schema upgrade on every deploy/reconfigure
manila-api manila-manage db sync (kolla docker/manila/manila-api/extend_start.sh:6-9) none bootstrap_manila, deploy/reconfigure and upgrade (roles/manila/tasks/bootstrap_service.yml:2-20, upgrade.yml:6) Schema sync on every deploy/reconfigure
nova-api nova-manage api_db sync, db sync, db online_data_migrations (kolla docker/nova/nova-api/extend_start.sh:9-14). Unreachable: Kolla-Ansible never sets it for Nova KOLLA_UPGRADE: api_db sync, db sync (:16-20). KOLLA_OSM: db online_data_migrations (:22-25) nova_api_bootstrap with KOLLA_UPGRADE, run_once, deploy/reconfigure and upgrade (roles/nova/tasks/bootstrap_service.yml:4-23, nova.yml:45-61), again when map_cell0 reports a change (roles/nova/tasks/map_cell0.yml:25-26). KOLLA_OSM, upgrade only (nova.yml:230-247) db sync targets cell0: the bootstrap nova.conf has [database] = cell0 (roles/nova/templates/nova.conf.j2:97-98). The bootstrap command is "false" (nova-api-bootstrap.json.j2:2), so success relies on the exit 0
nova-conductor nova-manage db sync --local_cell, db online_data_migrations (kolla docker/nova/nova-conductor/extend_start.sh:3-7). Unreachable KOLLA_UPGRADE: db sync --local_cell (:9-12). KOLLA_OSM: db online_data_migrations (:14-17) nova_cell_bootstrap with KOLLA_UPGRADE on the first conductor, deploy/reconfigure and upgrade (roles/nova-cell/tasks/bootstrap_service.yml:2-21, nova.yml:81-96). KOLLA_OSM, upgrade only (nova.yml:253-272) The same image runs nova_super_conductor (roles/nova/defaults/main.yml:80)
neutron-server neutron-db-manage with neutron.conf and ml2_conf.ini, upgrade head for neutron, neutron-dynamic-routing, neutron-fwaas and neutron-vpnaas, whether or not they are enabled (kolla docker/neutron/neutron-server/extend_start.sh:5-12) NEUTRON_SFC_BOOTSTRAP: networking-sfc upgrade head (:17-20). KOLLA_UPGRADE with NEUTRON_DB_EXPAND or NEUTRON_DB_CONTRACT: upgrade --expand or --contract for each NEUTRON_ROLLING_UPGRADE_SERVICES entry, without --config-file (:24-40) bootstrap_neutron on deploy, or on any action when neutron_enable_rolling_upgrade is off (default on) (roles/neutron/tasks/bootstrap_service.yml:2-22, roles/neutron/defaults/main.yml:411). bootstrap_neutron_sfc whenever enable_neutron_sfc (:24-46). Expand and contract containers in rolling_upgrade.yml:13-33,69-89 upgrade never migrates neutron-dynamic-routing (§6 item 8)
mariadb mysql_install_db, a temporary mysqld_safe --wsrep-new-cluster --skip-networking --wsrep-on=OFF, kolla_security_reset (root password), grants to root@localhost and root@'%', shutdown (kolla docker/mariadb/mariadb/extend_start.sh:5-26,37-41) KOLLA_UPGRADE: removes /var/lib/mysql/.pki, runs mysql_upgrade over TCP against the running local server (:44-54). BOOTSTRAP_ARGS: appended to mysqld_safe on any start (:56-58) bootstrap_mariadb when a new cluster is bootstrapped (roles/mariadb/tasks/bootstrap_cluster.yml:2-23), then a handler starts mariadb with BOOTSTRAP_ARGS=--wsrep-new-cluster (handlers/main.yml:2-20). upgrade_mariadb on every host on upgrade (tasks/upgrade.yml:4-26). --wsrep-recover and --wsrep-new-cluster in mariadb_recovery (recover_cluster.yml:41,122) The image forms no cluster itself and has no Galera recovery logic (§5.4)
rabbitmq writes RABBITMQ_CLUSTER_COOKIE to /var/lib/rabbitmq/.erlang.cookie, mode 400; nothing else: no rabbitmqctl, no join, no users (kolla docker/rabbitmq/extend_start.sh:7-18) none <project_name>_bootstrap, only when the volume was just created on that host (roles/rabbitmq/tasks/bootstrap.yml:10-26) A normal start never rewrites the cookie (§6 item 9)

No one-shot branch exists in: keystone-fernet, keystone-ssh, cinder-volume, cinder-scheduler, cinder-backup, heat-api-cfn, heat-engine, the barbican, octavia and manila worker images, all swift images, nova-scheduler, nova-compute, nova-compute-ironic, nova-libvirt, nova-ssh, the Nova console proxies, every neutron agent, openvswitch, ovn, mariadb-clustercheck, memcached, haproxy, keepalived, fluentd, cron, kolla-toolbox, chrony, iscsid and tgtd CODE. The only DB step in the OVS and OVN images is at start (§5).

Note on ironic (only this branch was read): ironic-api's KOLLA_BOOTSTRAP runs ironic-dbsync upgrade and online_data_migrations, KOLLA_UPGRADE only upgrade, KOLLA_OSM only online_data_migrations (kolla docker/ironic/ironic-api/extend_start.sh:5-19). bootstrap_ironic sends KOLLA_UPGRADE unless ironic_enable_rolling_upgrade (default yes) is turned off (roles/ironic/tasks/bootstrap_service.yml:2-25, roles/ironic/defaults/main.yml:206).

4.1 One-shots that bypass kolla_start

A few one-shots replace the container command with bash -c 'sudo -E kolla_set_configs && ...' CODE. The config is copied, but kolla_copy_cacerts and kolla_extend_start do not run, and the config.json command is written to /run_command but never executed. Custom CA certificates are therefore not installed in these containers.

Container Image What runs after kolla_set_configs Kolla-Ansible task
bootstrap_keystone_fernet keystone-fernet keystone-manage fernet_setup roles/keystone/tasks/bootstrap_service.yml:51-75
nova_api_map_cell0 nova-api nova-manage cell_v2 map_cell0 roles/nova/tasks/map_cell0.yml:2-23
nova_list_cells, cell create and update nova-conductor nova-manage cell_v2 list_cells, create_cell, update_cell roles/nova-cell/tasks/get_cell_settings.yml:8, create_cells.yml:12,38
mariabackup mariadb /usr/local/bin/kolla_mariadb_backup.sh (§5) roles/mariadb/tasks/backup.yml:1-22

The keystone-fernet image ships its own kolla_keystone_bootstrap, a fernet_setup wrapper, which Kolla-Ansible does not use. Its changed detection is inverted: it reports changed when "Key repository is already initialized" appears (kolla docker/keystone/keystone-fernet/keystone_bootstrap.sh:35-41). In this pairing that is dead code. discover_hosts is a docker exec into the running nova_conductor, not a one-shot (roles/nova-cell/tasks/discover_computes.yml:14-18).


5. Per-image start behaviour

"Runs as" is the image USER; root when the image and its parents set none. Every row starts with the common path of §2. "Log dir" means: create /var/log/kolla/<service> if missing and set mode 755.

Image Runs as What its start script does on every start Day-2 note
base, openstack-base root Nothing: empty kolla_extend_start (kolla docker/base/Dockerfile.j2:446) No USER in base, openstack-base or nova-base
keystone root Log dir and keystone.log with owners fixed, kolla_httpd_setup, ARGS=-DFOREGROUND (kolla docker/keystone/keystone/extend_start.sh:3-19,48-50). The command keystone-startup.sh waits up to 36 × 5 s for fernet key 0, then execs httpd (roles/keystone/templates/keystone-startup.sh.j2:9-24) Stale httpd run files are removed, so a restart is safe
keystone-fernet root Runs /usr/bin/fernet-node-sync.sh if present, then chowns the key directory (kolla docker/keystone/keystone-fernet/extend_start.sh:6-12). Kolla-Ansible installs that script; it waits up to 10 × 60 s for key 0, then exits 1 (roles/keystone/templates/fernet-node-sync.sh.j2:6-19) Without key 0 on the node the start waits up to 10 × 60 s and the script exits 1 before the cron command is exec'd (§2.2), so no rotation cron runs there meanwhile CODE; whether the container then restarts is NOT AVAILABLE
keystone-ssh root Generates any missing SSH host key (rsa, dsa, ecdsa, ed25519); creates /var/lib/keystone/.ssh (kolla docker/keystone/keystone-ssh/extend_start.sh:3-16) /etc/ssh is on no volume (roles/keystone/defaults/main.yml:37-42): new host keys on every recreate (§6 item 18)
glance-api glance Log dir (kolla docker/glance/glance-base/extend_start.sh:3-8); no httpd step. kolla_set_configs recursively chowns /var/lib/glance and /var/log/kolla/glance (roles/glance/templates/glance-api.json.j2:41-52) With the file backend /var/lib/glance is the image store: one chown per file on every start (§6 item 10)
placement-api root Log dir and placement-api.log; chown -R only if the top-level owner differs (kolla docker/placement/placement-base/extend_start.sh:3-13); kolla_httpd_setup (placement-api/extend_start.sh:23) –
cinder-api root Log dir (kolla docker/cinder/cinder-base/extend_start.sh:3-8); kolla_httpd_setup (cinder-api/extend_start.sh:20) –
cinder-volume cinder Log dir; sudo chown -R cinder:kolla of /var/lib/cinder and /etc/iscsi only if the top-level owner differs, chmod 2775 if needed (kolla docker/cinder/cinder-volume/extend_start.sh:3-14) Not only on a fresh volume: kolla_set_configs first chowns /var/lib/cinder recursively to cinder:cinder (roles/cinder/templates/cinder-volume.json.j2:38-43), so the cinder:kolla test always fails and chown -R walks the cinder volume a second time on every start
cinder-scheduler, cinder-backup cinder Log dir only (empty kolla_cinder_extend_start, kolla docker/cinder/cinder-base/Dockerfile.j2:78) –
heat-api, heat-api-cfn, heat-engine root (APIs), heat (engine) Log dir (kolla docker/heat/heat-base/extend_start.sh:3-8); the two API images run kolla_httpd_setup (heat-api/extend_start.sh:19, heat-api-cfn/extend_start.sh:3) –
horizon root Settings symlinks, plugin files from ENABLE_*, conditional static-asset regeneration, log dir, kolla_httpd_setup (kolla docker/horizon/extend_start.sh:24-42,51-388; §5.2) Assets are regenerated after a recreate or a settings change (§5.2); how long that start takes is NOT AVAILABLE
barbican, octavia and manila images barbican, octavia, manila (API images) Log dir (barbican also chowns it, kolla docker/barbican/barbican-base/extend_start.sh:4-13). The API images call kolla_httpd_setup, a no-op as a service user (barbican-api/Dockerfile.j2:48, octavia-api/Dockerfile.j2:78, manila-api/Dockerfile.j2:71) Worker images do the log-dir step only
swift-rsyncd swift sudo find /srv/node/ -maxdepth 1 -type d -execdir chown swift:swift {} + (not recursive) and mkdir -p /var/lib/swift/lock (kolla docker/swift/swift-rsyncd/extend_start.sh:3-4) Only rsyncd fixes the ownership of the disk mount points
swift account, container, object, object-expirer, proxy-server swift Nothing of their own Config and rings are copied only at container start, and the swift role notifies no handler (fact sheet D)
nova-base – Log dir, then sources kolla_nova_extend_start (kolla docker/nova/nova-base/extend_start.sh:3-10) Parent of every nova image except nova-libvirt
nova-api nova nova-base step; kolla_httpd_setup (kolla docker/nova/nova-api/extend_start.sh:27), a no-op as nova (Dockerfile.j2:71) No DB action on a normal start
nova-conductor, nova-scheduler nova nova-base step only Also nova_super_conductor. The scheduler SIGHUP comes from Kolla-Ansible (nova.yml:140-156)
nova-compute nova kolla_set_configs recursively chowns /var/log/kolla/nova and /var/lib/nova (roles/nova-cell/templates/nova-compute.json.j2:66-77); nova-base step; mkdir -p /var/lib/nova/instances (kolla docker/nova/nova-compute/extend_start.sh:3-5) /var/lib/nova is the instance datadir: one chown per file on every start (§6 item 10)
nova-compute-ironic nova nova-base step; permissions only on /var/log/kolla/nova (roles/nova-cell/templates/nova-compute-ironic.json.j2:19-21) –
nova-libvirt root /dev/kvm set to 660 root:qemu, xenfs mounted if loaded, log file created (kolla docker/nova/nova-libvirt/extend_start.sh:6-24); exec libvirtd --listen Passive towards running guests (§5.1)
nova-ssh root Replaces the nova-base script: generates missing SSH host keys, creates and chowns /var/lib/nova/.ssh (kolla docker/nova/nova-ssh/extend_start.sh:3-16); exec sshd -D New host keys on every recreate (§6 item 18)
nova-novncproxy, nova-serialproxy, nova-spicehtml5proxy nova nova-base step only –
neutron-base – Log dir; iptables backend: update-alternatives --set to the -legacy binaries when KOLLA_LEGACY_IPTABLES is "true", otherwise --auto (kolla docker/neutron/neutron-base/extend_start.sh:3-24) Only the OVS, linuxbridge and L3 agents get KOLLA_LEGACY_IPTABLES (roles/neutron/defaults/main.yml:32,67,93). Changing neutron_legacy_iptables recreates them; whether rules in the old backend are migrated is NOT AVAILABLE
neutron-server neutron neutron-base step; no branch taken; exec neutron-server –
neutron-openvswitch-agent neutron neutron-base step; nothing to OVS bridges or flows NOT AVAILABLE Whether and how the agent resyncs flows after a restart: agent behaviour, not in these sources
neutron-dhcp-agent neutron neutron-base step; no dnsmasq or namespace handling NOT AVAILABLE Whether dnsmasq children die with the container and whether qdhcp- namespaces persist on the shared /run/netns
neutron-l3-agent neutron neutron-base step; the command is Kolla-Ansible's wrapper (§5.3) Namespace cleanup on every start with enable_neutron_agent_ha (Risk register R6)
neutron-metadata-agent (also the OVN metadata agent) neutron neutron-base step; recursive chown of /var/lib/neutron/kolla, the socket volume shared with the L3 and DHCP agents (roles/neutron/templates/neutron-metadata-agent.json.j2:29-33) Ussuri runs the OVN metadata agent from this image (roles/neutron/defaults/main.yml:200)
neutron-sriov-agent, neutron-linuxbridge-agent neutron neutron-base step –
openvswitch-base root Log dir, then kolla_openvswitch_extend_start (kolla docker/openvswitch/openvswitch-base/extend_start.sh:3-10) –
openvswitch-db-server root mkdir -p /run/openvswitch; ovsdb-tool create if conf.db is absent (kolla docker/openvswitch/openvswitch-db-server/extend_start.sh:3-6). start-ovsdb-server converts the schema when needed, then runs ovsdb-server without exec (roles/openvswitch/templates/start-ovsdb-server.j2:5-11) conf.db is on the openvswitch_db volume: bridges and ports survive restart and recreate
openvswitch-vswitchd root modprobe openvswitch (kolla docker/openvswitch/openvswitch-vswitchd/extend_start.sh:3). start-ovs runs ovs-vswitchd without exec and without flow save or restore (roles/openvswitch/templates/start-ovs.j2:2) The kernel module stays loaded; whether flows are lost is NOT AVAILABLE (§6 item 13)
ovn-base root Replaces the OVS script: creates /run/ovn and the log dir (kolla docker/ovn/ovn-base/extend_start.sh:3-11) /run/ovn is on no volume
ovn-controller root ovn-base step; exec ovn-controller (roles/ovn/templates/ovn-controller.json.j2:2) Chassis settings are set by Kolla-Ansible with ovs-vsctl (roles/ovn/tasks/bootstrap.yml:2-45), not by the image
ovn-northd root ovn-base step; exec ovn-northd Stateless
ovn-nb-db-server, ovn-sb-db-server root ovn-base step. Kolla-Ansible calls /usr/share/ovn/scripts/ovn-ctl run_nb_ovsdb (or run_sb_ovsdb) directly, not the image's start-nb-db-server wrapper (roles/ovn/templates/ovn-nb-db.json.j2:2, ovn-sb-db.json.j2:2) The wrapper's comment says ovn-ctl creates the DB file when absent and updates it when the schema changes (kolla docker/ovn/ovn-nb-db-server/start_nb_db_server.sh:9-12); whether the direct call does so is NOT AVAILABLE
mariadb mysql kolla_set_configs recursively chowns /var/lib/mysql and /var/log/kolla/mariadb (roles/mariadb/templates/mariadb.json.j2:18-29); log dir; exec mysqld_safe plus BOOTSTRAP_ARGS if set (kolla docker/mariadb/mariadb/extend_start.sh:28-34,56-58) No Galera recovery in the image (§5.4)
mariadb (backup) mysql §4.1. backup.sh: mariabackup --backup --stream=xbstream, gzipped into /backup, full or incremental from BACKUP_TYPE (kolla docker/mariadb/mariadb/backup.sh:7-51) No retention or rotation in the image
mariadb-clustercheck root No start script. xinetd runs /usr/bin/clustercheck as mysql per connection (roles/mariadb/templates/mariadb-clustercheck.json.j2:2, xinetd-clustercheck.j2:9-12) Stateless. The clustercheck script is not in the kolla source
rabbitmq rabbitmq Recursive chown of /var/lib/rabbitmq and /var/log/kolla/<project_name> (roles/rabbitmq/templates/rabbitmq.json.j2:37-44); log dir (kolla docker/rabbitmq/extend_start.sh:20-26); cookie untouched Clustering is classic_config peer discovery (roles/rabbitmq/templates/rabbitmq.conf.j2:11-14), not the image
memcached memcached Nothing: no start script, empty config_files, all settings in the command (roles/memcached/templates/memcached.json.j2:2-3) No volume; whether a restart empties the cache is NOT AVAILABLE (§6 item 14)
haproxy root No start script. kolla_set_configs replaces haproxy.cfg and the whole services.d/ (roles/haproxy/templates/haproxy.json.j2:11-21); haproxy_run.sh starts haproxy -W -db (haproxy_run.sh.j2:7-9) The image's reload helper kolla_ensure_haproxy_latest_config (kolla docker/haproxy/ensure_latest_config.sh:5-14) is never called: config applies by restart only (Risk register R8)
keepalived root Removes a stale /run/keepalived.pid (kolla docker/keepalived/extend_start.sh:3-6); exec keepalived -nld VIP health is check_alive.sh, the haproxy stats socket only (§6 item 15)
fluentd td-agent or fluentd Sets owner and mode 2775 on /var/log/kolla, the kolla_logs volume every container shares, and chowns /var/lib/<binary>, through sudo (kolla docker/fluentd/extend_start.sh:9-28); recursive chowns of its log dirs and /var/lib/fluentd/data (roles/common/templates/fluentd.json.j2:26-58) NOT AVAILABLE Whether read positions under /var/run/<binary>/ are lost on recreate but kept on a restart
cron root No start script; copies logrotate.conf, exec crond -s -n or cron -f A restart only delays rotation
kolla-toolbox ansible No start script. When RabbitMQ is enabled, copies the Erlang cookie from passwords.yml (roles/common/templates/kolla-toolbox.json.j2:1-10); command sleep infinity Gets the current cookie on every start, unlike the brokers (§6 item 9)
chrony root Removes /var/run/chronyd.pid; log dir owned chrony:chrony (kolla docker/chrony/extend_start.sh:3-16) docker restart is safe
iscsid root Writes /etc/iscsi/initiatorname.iscsi with a new IQN only if the file is missing (kolla docker/iscsid/extend_start.sh:3-7) /etc/iscsi is the iscsi_info volume: the IQN is stable (§6 item 19)
tgtd root No start script, empty config_files; exec tgtd (roles/iscsi/templates/tgtd.json.j2:2-3) NOT AVAILABLE Whether the target table survives a restart, or exported targets drop until re-created

5.1 nova-libvirt: the start path is passive

The image is built from base, not nova-base (kolla docker/nova/nova-libvirt/Dockerfile.j2:1), and runs as root CODE. On every start kolla_set_configs deletes and re-copies only the files listed in roles/nova-cell/templates/nova-libvirt.json.j2:3-70 (libvirtd.conf, qemu.conf, /etc/libvirt/secrets, Ceph, SASL and TLS files); the manifest has no permissions block. extend_start.sh then changes /dev/kvm, which is the host's /dev, mounts xenfs only if lsmod shows it, and creates the log file (kolla docker/nova/nova-libvirt/extend_start.sh:6-24). Nothing touches cgroups, running QEMU processes, /var/run/libvirt, /var/lib/libvirt or /etc/libvirt/qemu, and no virtlogd or virtlockd is started. qemu.conf.j2:1-7 and libvirtd.conf.j2:1-23 set no cgroup option. Whether guests survive a restart is therefore decided by the stop path, where libvirt placed QEMU, and libvirtd's reconnect on start (Nova deep dive §5, Q1). The SASL DB /etc/libvirt/passwd.db is on no volume; only /etc/libvirt/qemu is (roles/nova-cell/defaults/main.yml:249). It survives a restart and is lost on every recreate, and the Create libvirt SASL user handler re-creates it (roles/nova-cell/handlers/main.yml:91,111-122).

5.2 horizon: static assets

On every start the script copies or removes each plugin's enabled/*.py file according to its ENABLE_* variable; any change sets FORCE_GENERATE=yes (kolla docker/horizon/extend_start.sh:51-65,339-366) CODE. It runs manage.py collectstatic --noinput --clear and compress --force only when /var/lib/kolla/.settings.md5sum.txt is missing, the md5 of local_settings, custom_local_settings and local_settings.d differs from it, or FORCE_GENERATE is yes; then it stores the new hash (:322-337,368-372). Kolla-Ansible sets FORCE_GENERATE=yes only with horizon_dev_mode (roles/horizon/defaults/main.yml:36). /var/lib/kolla is not a volume (:90-98), so the hash file lives in the container layer; only the start script writes it (HASH_PATH, kolla docker/horizon/extend_start.sh:6,371). A recreate (new image, changed parameters) or a settings change therefore makes the next start regenerate the assets; a plain restart with unchanged settings does not CODE. NOT AVAILABLE How long such a start takes.

5.3 neutron-l3-agent: the wrapper

The command is Kolla-Ansible's /usr/local/bin/neutron-l3-agent-wrapper.sh (roles/neutron/templates/neutron-l3-agent.json.j2:2-9) CODE. It sets errexit. With enable_neutron_agent_ha (default no) it first runs neutron-netns-cleanup --force --agent-type l3 (the wrapper's comment: "Remove all l3 related netns"); it then runs neutron-l3-agent without exec (neutron-l3-agent-wrapper.sh.j2:3-18). The cleanup runs on every container start: handler restart, recreate, Docker restart policy, reboot. A cleanup failure aborts the start. NOT AVAILABLE What the command deletes: it is neutron code, which is not in these sources (F12). Because the wrapper does not exec the agent, bash stays dumb-init's only child (§2.6) CODE. NOT AVAILABLE Whether docker stop's SIGTERM then reaches only bash, so that the agent gets no graceful shutdown, and whether its keepalived, haproxy and radvd children die with the container.

5.4 MariaDB: what the image does not do

The image has no Galera recovery logic: no grastate.dat or safe_to_bootstrap handling, no seqno comparison, no new-cluster decision CODE (kolla docker/mariadb/mariadb/extend_start.sh has none). All of that is in Kolla-Ansible (roles/mariadb/tasks/recover_cluster.yml, lookup_cluster.yml). A normal start is mysqld_safe with the rendered galera.cnf, whose gcomm:// lists every MariaDB host, or is empty on a single node (roles/mariadb/templates/galera.cnf.j2:29). BOOTSTRAP_ARGS is read from the container environment on every start (extend_start.sh:56-58), and the containers started with --wsrep-new-cluster have restart_policy: no (roles/mariadb/handlers/main.yml:2-20, recover_cluster.yml:115-127). The mariadb bootstrap, upgrade and --wsrep-recover one-shots also run the recursive /var/lib/mysql chown of §5, since they use the same manifest.


6. Day-2 consequences

  1. config.json-only changes written by genconfig never cause a restart. A normal deploy or reconfigure applies them, because the changed config.json template task notifies the restart handler (e.g. roles/memcached/tasks/config.yml:13-26). After a separate genconfig, --check ignores command and permissions, so deploy-containers sees no drift, and later runs do not restart for them either because the template task is then unchanged. Memcached is the extreme case. The change applies at the next restart for another reason CODE (Predict and inspect a run, Execution flow §4, Write config first, apply in controlled batches §2).
  2. Some containers restart on every deploy/reconfigure. A preserve_properties entry, such as prometheus extras, or a directory source with subdirectories makes --check exit 1 on every run (§3.1, §3.3) CODE (Tags §3).
  3. A new optional file fails the run. --check exits 2 when an optional file is rendered for the first time (e.g. nova-compute release) and no entry listed before it differs, and the check-containers task fails on that host instead of restarting CODE (Pre-execution checklist, Predict and inspect a run).
  4. A restart does not remove a dropped file. Destinations of entries removed from config.json stay in the container until it is recreated, and on a volume even after that (§3.5); --check does not report them. CA certificate files are purged and re-added on every start, but only from /etc/kolla/<svc>/ca-certificates/, where a CA deleted from kolla_certificates_dir/ca/ stays (the copy task never deletes, §3.5) and is re-added; whether the generated bundle keeps the last custom CA once it is removed there by hand is not known (§2.4) CODE (bundle and service effect NOT AVAILABLE) (Repository architecture §7).
  5. Nova's online data migrations never run outside upgrade. Deploy and reconfigure run the Nova bootstrap containers with KOLLA_UPGRADE (api_db sync + db sync, db sync --local_cell); the images' KOLLA_BOOTSTRAP branches are unreachable (§4) CODE (Nova deep dive §4, Findings F15).
  6. Placement's online data migrations run on every deploy/reconfigure, inside bootstrap_placement, and twice on upgrade. A tag change applied with deploy therefore runs them; cinder's bootstrap does not (§4) CODE (Update container images).
  7. A changed heat_domain_admin_password never reaches Keystone. The bootstrap only creates the user when heat_user_domain is absent, and a domain left without its user by a partial run is never repaired, while heat.conf gets the new value (roles/heat/templates/heat.conf.j2:12-13) CODE. NOT AVAILABLE Whether stack-domain operations then fail authentication (Known bugs and limitations).
  8. Neutron upgrade gaps. The default rolling upgrade never migrates neutron-dynamic-routing, which only the deploy bootstrap migrates. If has_offline_migrations returns non-zero for any subproject, or its docker exec fails (ignore_errors), every neutron_server is stopped until the final flush: a Neutron API outage (roles/neutron/tasks/rolling_upgrade.yml:43-93) CODE (Known bugs and limitations).
  9. The RabbitMQ cookie is fixed at bootstrap. The image writes .erlang.cookie only in the bootstrap container on a new volume. A changed rabbitmq_cluster_cookie recreates the brokers (env compare) but their cookie file keeps the old value, while kolla_toolbox gets the new one CODE (consequences NOT AVAILABLE) (fact sheet B §2).
  10. Recursive chown on data volumes at every start. kolla_set_configs makes one chown per file of the glance image store, of /var/lib/nova on computes, of the MariaDB datadir, of the RabbitMQ mnesia directory and of the cinder volume (twice for cinder-volume, §5), on every COPY_ALWAYS start CODE. NOT AVAILABLE Whether, with enable_shared_var_lib_nova_mnt, the walk also enters the shares still mounted under /var/lib/nova/mnt, and how start time grows with the number of files, especially on NFS-backed trees (fact sheet C, Nova deep dive §5).
  11. nova-libvirt's start path is passive; VM survival is decided on the stop path (§5.1) CODE (for the start path; NOT AVAILABLE for survival) (Nova deep dive §5, Q1).
  12. The L3 agent wrapper does not exec the agent. NOT AVAILABLE Whether SIGTERM then reaches only bash, so that the agent is killed without a graceful shutdown. With enable_neutron_agent_ha every start also runs neutron-netns-cleanup --force --agent-type l3 before the agent starts CODE; what that command deletes is NOT AVAILABLE (Risk register R6, Component dependencies §4).
  13. An openvswitch_vswitchd restart replaces only the userspace daemon. The kernel module stays loaded and conf.db keeps bridges and ports CODE. NOT AVAILABLE Whether the OpenFlow tables are lost and must be reprogrammed by the OVS agent or ovn-controller (Risk register R7, Q3).
  14. memcached has no data volume and no persistence option. The container mounts no data volume (roles/memcached/defaults/main.yml:37-40) and its command has no persistence option (roles/memcached/templates/memcached.json.j2:2) CODE. NOT AVAILABLE (memcached semantics) Whether every restart therefore empties the cache, with token-cache misses and lost Horizon sessions where they are kept in memcached (Risk register R4).
  15. The VIP follows the HAProxy stats socket only. keepalived's check_alive.sh asks the socket on the shared haproxy_socket volume (kolla docker/keepalived/check_alive.sh:6); with interval 2, fall 2, rise 10 (roles/haproxy/templates/keepalived.conf.j2:1-6) CODE. NOT AVAILABLE (keepalived semantics) Whether an HAProxy restart puts the node in FAULT, and for how long (Risk register R8).
  16. MariaDB has no automatic Galera recovery (§5.4). NOT AVAILABLE Whether, after all DB containers stopped (power loss, Docker restart on every DB host), they come back as plain mysqld_safe with the cluster non-primary until mariadb_recovery. A container left with BOOTSTRAP_ARGS=--wsrep-new-cluster by an interrupted run passes that option to mysqld_safe again on any manual docker start (§5.4) CODE; whether that bootstraps a new cluster (split-brain risk) is NOT AVAILABLE (fact sheet B §1, Risk register R18).
  17. No Docker HEALTHCHECK is wired in Ussuri (§2.7). docker ps shows no health status; service health must be checked from outside CODE (Pre-execution checklist item 5, health gates).
  18. SSH host keys of nova_ssh and keystone_ssh change on every recreate, because /etc/ssh is on no volume. The clients set StrictHostKeyChecking no and keep no known-hosts file (UserKnownHostsFile /dev/null, roles/nova-cell/templates/ssh_config.j2:2-3, roles/keystone/templates/ssh_config.j2:2-3) CODE; whether the SSH client then accepts the new keys without failing is NOT AVAILABLE (OpenSSH semantics) (Nova deep dive §6).
  19. The iSCSI initiator name is stable per iscsi_info volume across restart and recreate CODE. NOT AVAILABLE Whether removing the volume produces a new IQN at the next start (the start script writes one when the file is missing, §5), and whether that breaks backend ACLs keyed on the old one (Component dependencies §5).

7. What is not available from the sources

  • dumb-init signal forwarding. Whether --single-child forwards SIGTERM to the direct child only, and what happens to grandchildren (dnsmasq, keepalived and haproxy under the agents, services behind non-exec wrappers).
  • Whether the runtime kills the container's cgroup when its init exits, and how that applies under pid_mode: host (nova_libvirt).
  • QEMU survival across a nova_libvirt restart or recreate (Q1): placement in machine.slice through the host D-Bus and systemd-machined, reconnect from the runtime state, the old QEMU binary kept after a recreate (Nova deep dive §5).
  • Whether OVS flows are lost on an openvswitch_vswitchd restart and, if so, how long reprogramming takes (Q3).
  • Horizon regeneration time after a recreate or a settings change (§5.2).
  • Cost of the recursive chown on large trees: glance store, NFS instance directory, MariaDB datadir, RabbitMQ mnesia.
  • Upstream semantics of keystone-manage bootstrap: whether it resets the admin password to keystone_admin_password, and whether the wrapper, which counts nine "already exists"/"skipping" lines (kolla docker/keystone/keystone/keystone_bootstrap.sh:43-56), then still reports unchanged.
  • ovn-ctl schema creation and upgrade with Kolla-Ansible's direct call, and its hard-coded /usr/share/ovn/scripts/ path on distributions that ship ovn-ctl elsewhere.
  • /var/run as a symlink to /run in the images, which would put libvirt's runtime state on the host /run bind mount. The Dockerfiles do not show it.
  • Whether a failing kolla_set_configs causes a restart loop (§2.3) and the CA bundle after the last custom CA is removed (§2.4).
  • Service-side effects of the one-shot findings: heat stack-domain authentication after a changed heat_domain_admin_password (§6 item 7); whether rabbitmq-server ignores RABBITMQ_CLUSTER_COOKIE and uses only the cookie file (§6 item 9); Galera after a full stop, or a manual start left with --wsrep-new-cluster (§6 item 16).

To remember

Lumo: if you keep only 5 things from this page, keep these.

  • kolla_set_configs --check compares only the config_files entries of config.json (content, mode, owner). CODE
  • The check never compares command, permissions or removed entries.
  • A COPY_ALWAYS restart applies changed or added files, mode and owner changes and a changed command; removals do not apply. CODE
  • A newly rendered optional file whose destination is not yet in the container makes the check exit 2: the check task fails instead of restarting the container.
  • Run by hand, kolla_set_configs --check needs docker exec -u root.

Check yourself

Lumo: 3 quick questions to check the main ideas of this page. Choose an answer to see the explanation and the section that gives the evidence. Your score counts your first choices and stays in this browser only.

1. What does every container start run before the service command?
Show the answer

Answer: kolla_set_configs, kolla_copy_cacerts and the image's kolla_extend_start. kolla_start runs these three steps, then execs the service command. Read the section.

2. What does kolla_set_configs --check compare?
Show the answer

Answer: Only the config_files entries of config.json (content, mode, owner). It never compares command, permissions or removed entries CODE. Read the section.

3. A file is removed from the configuration of a COPY_ALWAYS container. Does a restart remove it?
Show the answer

Answer: No: removals do not apply. A COPY_ALWAYS restart applies changed or added files, mode and owner changes and a changed command CODE; removals do not apply. Read the section.