{ "manifest_version": "1.0.0", "template": { "id": "61571e20-995d-4204-b9b4-2522b4303eae", "slug": "Lm9gxI", "name": "LiteLLM", "description": "LiteLLM: LLM router, virtual keys, budgets, cost tracking, guardrails", "url": "https://railway.com/deploy/Lm9gxI", "upstream": { "image": "ghcr.io/berriai/litellm:v1.100.1@sha256:a3715fa7ad8387941ab697259bd2881d68931657247a41984f90fae6d11c62bf" } }, "services": [ { "name": "LiteLLM", "source": { "image": "ghcr.io/berriai/litellm:v1.100.1@sha256:a3715fa7ad8387941ab697259bd2881d68931657247a41984f90fae6d11c62bf" }, "needs_volume": false, "http": true }, { "name": "Postgres", "source": { "image": "ghcr.io/railwayapp-templates/postgres-ssl:18@sha256:469c779c7c57ec6bad4670a0a3cb5a830aa6e0ce4f3707137608de5223a5041c" }, "needs_volume": true, "volume_mount_path": "/var/lib/postgresql/data", "http": false }, { "name": "Redis", "source": { "image": "redis:8.2@sha256:7d1e4ce8b9395088377ab382d1f6cfdbd13b3690795198a0399ab8d683064d6d" }, "needs_volume": true, "volume_mount_path": "/data", "http": false } ], "required_inputs": [ { "key": "HOST", "service": "LiteLLM", "description": "Public listen address. Keep 0.0.0.0; the startup helper rejects a binding Railway cannot reach.", "secret": false, "strategy": "default", "default": "0.0.0.0" }, { "key": "PORT", "service": "LiteLLM", "description": "Port used by both LiteLLM and Railway readiness. Default 4000; the startup helper validates and passes this value to the server.", "secret": false, "strategy": "default", "default": "4000" }, { "key": "REDIS_URL", "service": "LiteLLM", "description": "Enables response caching and coordinates rate limits across replicas.", "secret": false, "strategy": "railway_provided", "railway_source": "reference_variable" }, { "key": "OR_SITE_URL", "service": "LiteLLM", "description": "Optional - Sent to OpenRouter as HTTP-Referer for attribution.", "secret": false, "strategy": "railway_provided", "railway_source": "railway_domain" }, { "key": "UI_USERNAME", "service": "LiteLLM", "description": "Optional - Admin UI username.", "secret": false, "strategy": "default", "default": "admin" }, { "key": "DATABASE_URL", "service": "LiteLLM", "description": "Stores models, virtual keys, budgets and spend.", "secret": false, "strategy": "railway_provided", "railway_source": "reference_variable" }, { "key": "LITELLM_SALT_KEY", "service": "LiteLLM", "description": "Encrypts provider credentials at rest. Generated once and NEVER changed - rotating it silently makes every stored provider key unreadable.", "secret": true, "strategy": "generate", "generate": "random_base64_32" }, { "key": "STORE_MODEL_IN_DB", "service": "LiteLLM", "description": "Lets you add models from the Admin UI. Without it, Add Model returns HTTP 500.", "secret": false, "strategy": "default", "default": "True" }, { "key": "LITELLM_MASTER_KEY", "service": "LiteLLM", "description": "Your API key AND the Admin UI password. Generated for you - copy it from this service's Variables tab after deploying.", "secret": true, "strategy": "generate", "generate": "random_base64_32" }, { "key": "LITELLM_LOCAL_MODEL_COST_MAP", "service": "LiteLLM", "description": "Uses the bundled pricing table instead of fetching it at boot. Halves startup time.", "secret": false, "strategy": "default", "default": "True" }, { "key": "ENFORCE_PRISMA_MIGRATION_CHECK", "service": "LiteLLM", "description": "Fails the deploy loudly on a bad migration instead of serving a half-migrated database.", "secret": false, "strategy": "default", "default": "true" }, { "key": "RAILWAY_DEPLOYMENT_DRAINING_SECONDS", "service": "LiteLLM", "description": "Allows 120 seconds for the previous gateway deployment to drain. Long requests may exceed this; dependency redeploys still interrupt service.", "secret": false, "strategy": "default", "default": "120" }, { "key": "PGDATA", "service": "Postgres", "description": "Postgres data directory, inside the mounted volume.", "secret": false, "strategy": "default", "default": "/var/lib/postgresql/data/pgdata" }, { "key": "PGHOST", "service": "Postgres", "description": "Private-network hostname of the Postgres service.", "secret": false, "strategy": "railway_provided", "railway_source": "railway_private_domain" }, { "key": "PGPORT", "service": "Postgres", "description": "Port Postgres listens on.", "secret": false, "strategy": "default", "default": "5432" }, { "key": "PGUSER", "service": "Postgres", "description": "Postgres superuser name.", "secret": false, "strategy": "railway_provided", "railway_source": "reference_variable" }, { "key": "PGDATABASE", "service": "Postgres", "description": "Default database name used by psql and the Postgres client tools.", "secret": false, "strategy": "railway_provided", "railway_source": "reference_variable" }, { "key": "PGPASSWORD", "service": "Postgres", "description": "Password for the Postgres superuser.", "secret": true, "strategy": "railway_provided", "railway_source": "reference_variable" }, { "key": "POSTGRES_DB", "service": "Postgres", "description": "Database created on first boot.", "secret": false, "strategy": "default", "default": "railway" }, { "key": "DATABASE_URL", "service": "Postgres", "description": "Private-network connection string. LiteLLM stores models, virtual keys, budgets and spend history here.", "secret": false, "strategy": "railway_provided", "railway_source": "railway_private_domain" }, { "key": "POSTGRES_USER", "service": "Postgres", "description": "Superuser created on first boot.", "secret": false, "strategy": "default", "default": "postgres" }, { "key": "SSL_CERT_DAYS", "service": "Postgres", "description": "Validity window, in days, for the self-signed TLS certificate this image generates.", "secret": false, "strategy": "default", "default": "820" }, { "key": "POSTGRES_PASSWORD", "service": "Postgres", "description": "Superuser password, generated once at deploy time.", "secret": true, "strategy": "generate", "generate": "strong_password" }, { "key": "RAILWAY_DEPLOYMENT_DRAINING_SECONDS", "service": "Postgres", "description": "Seconds Railway keeps the old container alive after the new one is live, so in-flight requests can finish.", "secret": false, "strategy": "default", "default": "60" }, { "key": "REDISHOST", "service": "Redis", "description": "Private-network hostname of the Redis service.", "secret": false, "strategy": "railway_provided", "railway_source": "railway_private_domain" }, { "key": "REDISPORT", "service": "Redis", "description": "Port Redis listens on.", "secret": false, "strategy": "default", "default": "6379" }, { "key": "REDISUSER", "service": "Redis", "description": "Redis username for this image's default user.", "secret": false, "strategy": "default", "default": "default" }, { "key": "REDIS_URL", "service": "Redis", "description": "Private-network connection string. LiteLLM uses Redis for response caching and cross-replica rate limiting.", "secret": false, "strategy": "railway_provided", "railway_source": "railway_private_domain" }, { "key": "REDIS_PASSWORD", "service": "Redis", "description": "Redis password, generated once at deploy time.", "secret": true, "strategy": "generate", "generate": "strong_password" }, { "key": "RAILWAY_DEPLOYMENT_DRAINING_SECONDS", "service": "Redis", "description": "Seconds Railway keeps the old container alive after the new one is live, so in-flight requests can finish.", "secret": false, "strategy": "default", "default": "60" } ], "deploy": { "mcp": { "server": "railway", "tool": "deploy_template", "args": { "template_code": "Lm9gxI" } }, "cli": "railway deploy --template Lm9gxI", "api": { "method": "POST", "path": "/graphql/v2", "body": { "query": "mutation templateDeploy($input: TemplateDeployV2Input!) { templateDeployV2(input: $input) { projectId workflowId } }", "variables": { "input": { "templateId": "61571e20-995d-4204-b9b4-2522b4303eae", "serializedConfig": { "services": { "00af68c0-dab6-4206-9b37-75db9a6f3b73": { "icon": "https://bogusz.co/external/litellm-icon.png", "name": "LiteLLM", "deploy": { "startCommand": "python -c 'exec('\"'\"'\"\"\"Embedded startup command for the official LiteLLM image (no LiteLLM imports).\\n\\npredeploy: validate, wait for Postgres, run `prisma migrate deploy` directly (~90 MiB).\\nserve: validate, wait for Postgres, exec litellm with the Prisma CLI hidden so the\\n proxy skips its own startup migration pass (~500 MiB of Node on top of Python).\\n\"\"\"\\n\\nimport importlib.util\\nimport os\\nimport re\\nimport signal\\nimport subprocess\\nimport sys\\nimport tempfile\\nimport time\\n\\nimport yaml\\n\\nMIB = 2**20\\n# Measured on the pinned image: serve peaks ~800 MiB at readiness, ~1.3 GiB with LiteLLM\\'\"'\"'s own\\n# startup migration pass. Below MIN_MEMORY nothing works; below COMFORTABLE_MEMORY it runs but\\n# has little headroom under load.\\nMIN_MEMORY = 900 * MIB\\nCOMFORTABLE_MEMORY = 1536 * MIB\\n\\n\\nclass StartupError(Exception):\\n \"\"\"Safe-to-log startup failure; never include user values or driver messages.\"\"\"\\n\\n\\ndef validate_environment(environ):\\n for name in (\"LITELLM_MASTER_KEY\", \"LITELLM_SALT_KEY\"):\\n if not environ.get(name, \"\").strip():\\n raise StartupError(f\"{name} must be nonblank\")\\n port = environ.get(\"PORT\", \"4000\")\\n if not re.fullmatch(r\"[0-9]{1,5}\", port) or not 1 <= int(port) <= 65535:\\n raise StartupError(\"PORT must be an integer from 1 to 65535\")\\n host = environ.get(\"HOST\", \"0.0.0.0\")\\n if host != \"0.0.0.0\":\\n raise StartupError(\"HOST must be 0.0.0.0 for the template public binding\")\\n return host, str(int(port))\\n\\n\\ndef validate_config(raw):\\n if not raw or not raw.strip():\\n return None\\n try:\\n config = yaml.safe_load(raw)\\n except yaml.YAMLError as exc:\\n mark = getattr(exc, \"problem_mark\", None)\\n location = f\" at line {mark.line + 1}, column {mark.column + 1}\" if mark else \"\"\\n raise StartupError(f\"LITELLM_CONFIG_YAML is invalid YAML{location}\") from None\\n if not isinstance(config, dict) or not config:\\n raise StartupError(\"LITELLM_CONFIG_YAML root must be a nonempty mapping\")\\n for key in (\"general_settings\", \"litellm_settings\", \"router_settings\"):\\n if key in config and not isinstance(config[key], dict):\\n raise StartupError(f\"LITELLM_CONFIG_YAML {key} must be a mapping\")\\n if \"model_list\" in config and not isinstance(config[\"model_list\"], list):\\n raise StartupError(\"LITELLM_CONFIG_YAML model_list must be a list\")\\n return config\\n\\n\\ndef resolve_database_url(config, environ):\\n \"\"\"Match LiteLLM\\'\"'\"'s YAML-over-environment database selection before probing.\"\"\"\\n configured = (config or {}).get(\"general_settings\", {}).get(\"database_url\")\\n value = configured if configured is not None else environ.get(\"DATABASE_URL\")\\n if isinstance(value, str) and value.startswith(\"os.environ/\"):\\n value = environ.get(value.removeprefix(\"os.environ/\"))\\n if not isinstance(value, str) or not value.strip():\\n raise StartupError(\"Database URL must be nonblank; check DATABASE_URL and general_settings.database_url\")\\n return value\\n\\n\\ndef memory_limit(path=\"/sys/fs/cgroup/memory.max\"):\\n \"\"\"Container memory limit in bytes, or None when unlimited/unknown.\"\"\"\\n try:\\n with open(path) as f:\\n value = f.read().strip()\\n except OSError:\\n return None\\n return int(value) if value.isdigit() else None\\n\\n\\ndef check_memory(limit):\\n if limit is None:\\n return\\n if limit < MIN_MEMORY:\\n raise StartupError(\\n f\"this container allows {limit // MIB} MiB of memory; LiteLLM needs at least 1 GB to start \"\\n \"(Railway Trial) and 2 GB or more for real use (Hobby plan or higher). Upgrade the plan or \"\\n \"raise the service memory limit, then redeploy\"\\n )\\n if limit < COMFORTABLE_MEMORY:\\n print(\\n f\"LiteLLM startup: {limit // MIB} MiB memory limit is enough to start but leaves little headroom; \"\\n \"expect restarts under load. 2 GB or more is recommended\",\\n file=sys.stderr, flush=True,\\n )\\n\\n\\ndef prisma_cli():\\n cli = os.environ.get(\"PRISMA_CLI_PATH\")\\n if not cli:\\n raise StartupError(\"The pinned image must provide PRISMA_CLI_PATH\")\\n return cli\\n\\n\\ndef database_probe(dsn, timeout):\\n \"\"\"Use the image\\'\"'\"'s offline Prisma toolchain, not an extra Python DB driver.\"\"\"\\n command = [prisma_cli(), \"db\", \"execute\", \"--url\", dsn, \"--stdin\"]\\n with subprocess.Popen(\\n command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE,\\n text=True, start_new_session=True,\\n ) as process:\\n try:\\n _, error = process.communicate(\"SELECT 1;\", timeout=timeout)\\n except subprocess.TimeoutExpired:\\n # Retire the native schema-engine child as well as the CLI parent.\\n os.killpg(process.pid, signal.SIGKILL)\\n process.communicate()\\n return False\\n if process.returncode == 0:\\n return True\\n # Driver output can contain credentials and SQL. Classify it, never print it.\\n code = re.search(r\"\\\\bP[0-9]{4}\\\\b\", error)\\n if code and code.group() in {\"P1001\", \"P1002\", \"P1008\", \"P1017\", \"P2024\", \"P2037\"}:\\n return False\\n if \"database system is starting up\" in error.lower():\\n return False\\n raise StartupError(\"Database connection rejected; check DATABASE_URL and credentials\")\\n\\n\\ndef wait_for_database(dsn):\\n deadline = time.monotonic() + 90\\n waiting = False\\n while (remaining := deadline - time.monotonic()) > 0:\\n if database_probe(dsn, min(5, remaining)):\\n return\\n if not waiting:\\n print(\"LiteLLM startup: waiting up to 90 seconds for database readiness\", file=sys.stderr, flush=True)\\n waiting = True\\n remaining = deadline - time.monotonic()\\n if remaining > 0:\\n time.sleep(min(2, remaining))\\n raise StartupError(\"Database readiness timed out after 90 seconds\")\\n\\n\\ndef migrations_dir():\\n \"\"\"Directory holding LiteLLM\\'\"'\"'s schema.prisma and migrations, found without importing it.\"\"\"\\n spec = importlib.util.find_spec(\"litellm_proxy_extras\")\\n if spec is None or not spec.submodule_search_locations:\\n raise StartupError(\"The pinned image must provide litellm_proxy_extras\")\\n return list(spec.submodule_search_locations)[0]\\n\\n\\ndef migrate(dsn):\\n \"\"\"`prisma migrate deploy` exactly as LiteLLM\\'\"'\"'s v2 resolver runs it, minus a 1.3 GiB Python import.\"\"\"\\n env = os.environ | {\"DATABASE_URL\": dsn, \"NPM_CONFIG_PREFER_OFFLINE\": \"true\", \"PRISMA_HIDE_UPDATE_MESSAGE\": \"1\"}\\n result = subprocess.run([prisma_cli(), \"migrate\", \"deploy\"], cwd=migrations_dir(), env=env)\\n if result.returncode:\\n raise StartupError(\"database migration failed; see the Prisma output above\")\\n\\n\\ndef serve_path(environ):\\n \"\"\"PATH with every venv tool except `prisma`: LiteLLM then skips its startup migration pass.\"\"\"\\n venv_bin = os.path.dirname(sys.executable)\\n shim = tempfile.mkdtemp(prefix=\"litellm-bin-\")\\n for name in os.listdir(venv_bin):\\n if name != \"prisma\":\\n os.symlink(os.path.join(venv_bin, name), os.path.join(shim, name))\\n rest = [p for p in environ.get(\"PATH\", \"\").split(os.pathsep) if p and p != venv_bin]\\n return os.pathsep.join([shim, *rest])\\n\\n\\ndef fail(message):\\n \"\"\"Print the reason, then linger briefly: Railway drops the logs of containers that exit within ~1 s.\"\"\"\\n print(message, file=sys.stderr, flush=True)\\n time.sleep(3)\\n return 1\\n\\n\\ndef main():\\n path = None\\n try:\\n if len(sys.argv) != 2 or sys.argv[1] not in (\"predeploy\", \"serve\"):\\n raise StartupError(\"Expected mode: predeploy or serve\")\\n host, port = validate_environment(os.environ)\\n raw = os.environ.get(\"LITELLM_CONFIG_YAML\")\\n parsed = validate_config(raw)\\n check_memory(memory_limit())\\n dsn = resolve_database_url(parsed, os.environ)\\n wait_for_database(dsn)\\n if sys.argv[1] == \"predeploy\":\\n migrate(dsn)\\n return 0\\n command = [os.path.join(os.path.dirname(sys.executable), \"litellm\"), \"--host\", host, \"--port\", port]\\n if parsed is not None:\\n # NamedTemporaryFile creates mode 0600. Preserve the original YAML,\\n # including environment references, rather than merging or dumping it.\\n with tempfile.NamedTemporaryFile(mode=\"w\", encoding=\"utf-8\", suffix=\".yaml\", delete=False) as config:\\n path = config.name\\n config.write(raw)\\n command.extend([\"--config\", path])\\n os.execve(command[0], command, os.environ | {\"PATH\": serve_path(os.environ)})\\n except StartupError as exc:\\n return fail(f\"LiteLLM startup: {exc}\")\\n except OSError:\\n return fail(\"LiteLLM startup: unable to write config or execute litellm\")\\n finally:\\n # Successful exec leaves the private file available for LiteLLM to read.\\n if path is not None:\\n os.unlink(path)\\n\\n\\nif __name__ == \"__main__\":\\n sys.exit(main())\\n'\"'\"')' serve", "healthcheckPath": "/health/readiness", "preDeployCommand": [ "python -c 'exec('\"'\"'\"\"\"Embedded startup command for the official LiteLLM image (no LiteLLM imports).\\n\\npredeploy: validate, wait for Postgres, run `prisma migrate deploy` directly (~90 MiB).\\nserve: validate, wait for Postgres, exec litellm with the Prisma CLI hidden so the\\n proxy skips its own startup migration pass (~500 MiB of Node on top of Python).\\n\"\"\"\\n\\nimport importlib.util\\nimport os\\nimport re\\nimport signal\\nimport subprocess\\nimport sys\\nimport tempfile\\nimport time\\n\\nimport yaml\\n\\nMIB = 2**20\\n# Measured on the pinned image: serve peaks ~800 MiB at readiness, ~1.3 GiB with LiteLLM\\'\"'\"'s own\\n# startup migration pass. Below MIN_MEMORY nothing works; below COMFORTABLE_MEMORY it runs but\\n# has little headroom under load.\\nMIN_MEMORY = 900 * MIB\\nCOMFORTABLE_MEMORY = 1536 * MIB\\n\\n\\nclass StartupError(Exception):\\n \"\"\"Safe-to-log startup failure; never include user values or driver messages.\"\"\"\\n\\n\\ndef validate_environment(environ):\\n for name in (\"LITELLM_MASTER_KEY\", \"LITELLM_SALT_KEY\"):\\n if not environ.get(name, \"\").strip():\\n raise StartupError(f\"{name} must be nonblank\")\\n port = environ.get(\"PORT\", \"4000\")\\n if not re.fullmatch(r\"[0-9]{1,5}\", port) or not 1 <= int(port) <= 65535:\\n raise StartupError(\"PORT must be an integer from 1 to 65535\")\\n host = environ.get(\"HOST\", \"0.0.0.0\")\\n if host != \"0.0.0.0\":\\n raise StartupError(\"HOST must be 0.0.0.0 for the template public binding\")\\n return host, str(int(port))\\n\\n\\ndef validate_config(raw):\\n if not raw or not raw.strip():\\n return None\\n try:\\n config = yaml.safe_load(raw)\\n except yaml.YAMLError as exc:\\n mark = getattr(exc, \"problem_mark\", None)\\n location = f\" at line {mark.line + 1}, column {mark.column + 1}\" if mark else \"\"\\n raise StartupError(f\"LITELLM_CONFIG_YAML is invalid YAML{location}\") from None\\n if not isinstance(config, dict) or not config:\\n raise StartupError(\"LITELLM_CONFIG_YAML root must be a nonempty mapping\")\\n for key in (\"general_settings\", \"litellm_settings\", \"router_settings\"):\\n if key in config and not isinstance(config[key], dict):\\n raise StartupError(f\"LITELLM_CONFIG_YAML {key} must be a mapping\")\\n if \"model_list\" in config and not isinstance(config[\"model_list\"], list):\\n raise StartupError(\"LITELLM_CONFIG_YAML model_list must be a list\")\\n return config\\n\\n\\ndef resolve_database_url(config, environ):\\n \"\"\"Match LiteLLM\\'\"'\"'s YAML-over-environment database selection before probing.\"\"\"\\n configured = (config or {}).get(\"general_settings\", {}).get(\"database_url\")\\n value = configured if configured is not None else environ.get(\"DATABASE_URL\")\\n if isinstance(value, str) and value.startswith(\"os.environ/\"):\\n value = environ.get(value.removeprefix(\"os.environ/\"))\\n if not isinstance(value, str) or not value.strip():\\n raise StartupError(\"Database URL must be nonblank; check DATABASE_URL and general_settings.database_url\")\\n return value\\n\\n\\ndef memory_limit(path=\"/sys/fs/cgroup/memory.max\"):\\n \"\"\"Container memory limit in bytes, or None when unlimited/unknown.\"\"\"\\n try:\\n with open(path) as f:\\n value = f.read().strip()\\n except OSError:\\n return None\\n return int(value) if value.isdigit() else None\\n\\n\\ndef check_memory(limit):\\n if limit is None:\\n return\\n if limit < MIN_MEMORY:\\n raise StartupError(\\n f\"this container allows {limit // MIB} MiB of memory; LiteLLM needs at least 1 GB to start \"\\n \"(Railway Trial) and 2 GB or more for real use (Hobby plan or higher). Upgrade the plan or \"\\n \"raise the service memory limit, then redeploy\"\\n )\\n if limit < COMFORTABLE_MEMORY:\\n print(\\n f\"LiteLLM startup: {limit // MIB} MiB memory limit is enough to start but leaves little headroom; \"\\n \"expect restarts under load. 2 GB or more is recommended\",\\n file=sys.stderr, flush=True,\\n )\\n\\n\\ndef prisma_cli():\\n cli = os.environ.get(\"PRISMA_CLI_PATH\")\\n if not cli:\\n raise StartupError(\"The pinned image must provide PRISMA_CLI_PATH\")\\n return cli\\n\\n\\ndef database_probe(dsn, timeout):\\n \"\"\"Use the image\\'\"'\"'s offline Prisma toolchain, not an extra Python DB driver.\"\"\"\\n command = [prisma_cli(), \"db\", \"execute\", \"--url\", dsn, \"--stdin\"]\\n with subprocess.Popen(\\n command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE,\\n text=True, start_new_session=True,\\n ) as process:\\n try:\\n _, error = process.communicate(\"SELECT 1;\", timeout=timeout)\\n except subprocess.TimeoutExpired:\\n # Retire the native schema-engine child as well as the CLI parent.\\n os.killpg(process.pid, signal.SIGKILL)\\n process.communicate()\\n return False\\n if process.returncode == 0:\\n return True\\n # Driver output can contain credentials and SQL. Classify it, never print it.\\n code = re.search(r\"\\\\bP[0-9]{4}\\\\b\", error)\\n if code and code.group() in {\"P1001\", \"P1002\", \"P1008\", \"P1017\", \"P2024\", \"P2037\"}:\\n return False\\n if \"database system is starting up\" in error.lower():\\n return False\\n raise StartupError(\"Database connection rejected; check DATABASE_URL and credentials\")\\n\\n\\ndef wait_for_database(dsn):\\n deadline = time.monotonic() + 90\\n waiting = False\\n while (remaining := deadline - time.monotonic()) > 0:\\n if database_probe(dsn, min(5, remaining)):\\n return\\n if not waiting:\\n print(\"LiteLLM startup: waiting up to 90 seconds for database readiness\", file=sys.stderr, flush=True)\\n waiting = True\\n remaining = deadline - time.monotonic()\\n if remaining > 0:\\n time.sleep(min(2, remaining))\\n raise StartupError(\"Database readiness timed out after 90 seconds\")\\n\\n\\ndef migrations_dir():\\n \"\"\"Directory holding LiteLLM\\'\"'\"'s schema.prisma and migrations, found without importing it.\"\"\"\\n spec = importlib.util.find_spec(\"litellm_proxy_extras\")\\n if spec is None or not spec.submodule_search_locations:\\n raise StartupError(\"The pinned image must provide litellm_proxy_extras\")\\n return list(spec.submodule_search_locations)[0]\\n\\n\\ndef migrate(dsn):\\n \"\"\"`prisma migrate deploy` exactly as LiteLLM\\'\"'\"'s v2 resolver runs it, minus a 1.3 GiB Python import.\"\"\"\\n env = os.environ | {\"DATABASE_URL\": dsn, \"NPM_CONFIG_PREFER_OFFLINE\": \"true\", \"PRISMA_HIDE_UPDATE_MESSAGE\": \"1\"}\\n result = subprocess.run([prisma_cli(), \"migrate\", \"deploy\"], cwd=migrations_dir(), env=env)\\n if result.returncode:\\n raise StartupError(\"database migration failed; see the Prisma output above\")\\n\\n\\ndef serve_path(environ):\\n \"\"\"PATH with every venv tool except `prisma`: LiteLLM then skips its startup migration pass.\"\"\"\\n venv_bin = os.path.dirname(sys.executable)\\n shim = tempfile.mkdtemp(prefix=\"litellm-bin-\")\\n for name in os.listdir(venv_bin):\\n if name != \"prisma\":\\n os.symlink(os.path.join(venv_bin, name), os.path.join(shim, name))\\n rest = [p for p in environ.get(\"PATH\", \"\").split(os.pathsep) if p and p != venv_bin]\\n return os.pathsep.join([shim, *rest])\\n\\n\\ndef fail(message):\\n \"\"\"Print the reason, then linger briefly: Railway drops the logs of containers that exit within ~1 s.\"\"\"\\n print(message, file=sys.stderr, flush=True)\\n time.sleep(3)\\n return 1\\n\\n\\ndef main():\\n path = None\\n try:\\n if len(sys.argv) != 2 or sys.argv[1] not in (\"predeploy\", \"serve\"):\\n raise StartupError(\"Expected mode: predeploy or serve\")\\n host, port = validate_environment(os.environ)\\n raw = os.environ.get(\"LITELLM_CONFIG_YAML\")\\n parsed = validate_config(raw)\\n check_memory(memory_limit())\\n dsn = resolve_database_url(parsed, os.environ)\\n wait_for_database(dsn)\\n if sys.argv[1] == \"predeploy\":\\n migrate(dsn)\\n return 0\\n command = [os.path.join(os.path.dirname(sys.executable), \"litellm\"), \"--host\", host, \"--port\", port]\\n if parsed is not None:\\n # NamedTemporaryFile creates mode 0600. Preserve the original YAML,\\n # including environment references, rather than merging or dumping it.\\n with tempfile.NamedTemporaryFile(mode=\"w\", encoding=\"utf-8\", suffix=\".yaml\", delete=False) as config:\\n path = config.name\\n config.write(raw)\\n command.extend([\"--config\", path])\\n os.execve(command[0], command, os.environ | {\"PATH\": serve_path(os.environ)})\\n except StartupError as exc:\\n return fail(f\"LiteLLM startup: {exc}\")\\n except OSError:\\n return fail(\"LiteLLM startup: unable to write config or execute litellm\")\\n finally:\\n # Successful exec leaves the private file available for LiteLLM to read.\\n if path is not None:\\n os.unlink(path)\\n\\n\\nif __name__ == \"__main__\":\\n sys.exit(main())\\n'\"'\"')' predeploy" ] }, "source": { "image": "ghcr.io/berriai/litellm:v1.100.1@sha256:a3715fa7ad8387941ab697259bd2881d68931657247a41984f90fae6d11c62bf" }, "variables": { "HOST": { "isOptional": false, "description": "Public listen address. Keep 0.0.0.0; the startup helper rejects a binding Railway cannot reach.", "defaultValue": "0.0.0.0" }, "PORT": { "isOptional": false, "description": "Port used by both LiteLLM and Railway readiness. Default 4000; the startup helper validates and passes this value to the server.", "defaultValue": "4000" }, "REDIS_URL": { "description": "Enables response caching and coordinates rate limits across replicas.", "defaultValue": "${{Redis.REDIS_URL}}" }, "OR_APP_NAME": { "isOptional": true, "description": "Optional - Sent to OpenRouter as X-Title for attribution.", "defaultValue": "" }, "OR_SITE_URL": { "isOptional": true, "description": "Optional - Sent to OpenRouter as HTTP-Referer for attribution.", "defaultValue": "https://${{RAILWAY_PUBLIC_DOMAIN}}" }, "UI_PASSWORD": { "isOptional": true, "description": "Optional - Leave blank to log in with LITELLM_MASTER_KEY.", "defaultValue": "" }, "UI_USERNAME": { "isOptional": true, "description": "Optional - Admin UI username.", "defaultValue": "admin" }, "DATABASE_URL": { "isOptional": false, "description": "Stores models, virtual keys, budgets and spend.", "defaultValue": "${{Postgres.DATABASE_URL}}" }, "GEMINI_API_KEY": { "isOptional": true, "description": "Optional", "defaultValue": "" }, "OPENAI_API_KEY": { "isOptional": true, "description": "Optional - Models are normally added from the Admin UI instead.", "defaultValue": "" }, "LITELLM_SALT_KEY": { "isOptional": false, "description": "Encrypts provider credentials at rest. Generated once and NEVER changed - rotating it silently makes every stored provider key unreadable.", "defaultValue": "{{LITELLM_SALT_KEY}}" }, "ANTHROPIC_API_KEY": { "isOptional": true, "description": "Optional", "defaultValue": "" }, "STORE_MODEL_IN_DB": { "isOptional": false, "description": "Lets you add models from the Admin UI. Without it, Add Model returns HTTP 500.", "defaultValue": "True" }, "LITELLM_MASTER_KEY": { "isOptional": false, "description": "Your API key AND the Admin UI password. Generated for you - copy it from this service's Variables tab after deploying.", "defaultValue": "{{LITELLM_MASTER_KEY}}" }, "OPENROUTER_API_KEY": { "isOptional": true, "description": "Optional - One key unlocks ~100 models via the openrouter/* wildcard.", "defaultValue": "" }, "LITELLM_CONFIG_YAML": { "isOptional": true, "description": "Optional - Advanced config.yaml, e.g. Prometheus. Whitespace-only is ignored; invalid YAML fails pre-deploy before migrations. Max 32768 characters.", "defaultValue": "" }, "LITELLM_LOCAL_MODEL_COST_MAP": { "description": "Uses the bundled pricing table instead of fetching it at boot. Halves startup time.", "defaultValue": "True" }, "ENFORCE_PRISMA_MIGRATION_CHECK": { "description": "Fails the deploy loudly on a bad migration instead of serving a half-migrated database.", "defaultValue": "true" }, "RAILWAY_DEPLOYMENT_DRAINING_SECONDS": { "description": "Allows 120 seconds for the previous gateway deployment to drain. Long requests may exceed this; dependency redeploys still interrupt service.", "defaultValue": "120" } }, "networking": { "serviceDomains": { "": {} } } }, "6d63359f-deb4-46e7-a1c8-08de18538182": { "icon": "https://devicons.railway.app/i/postgresql.svg", "name": "Postgres", "deploy": { "startCommand": null, "healthcheckPath": null }, "source": { "image": "ghcr.io/railwayapp-templates/postgres-ssl:18@sha256:469c779c7c57ec6bad4670a0a3cb5a830aa6e0ce4f3707137608de5223a5041c" }, "variables": { "PGDATA": { "isOptional": false, "description": "Postgres data directory, inside the mounted volume.", "defaultValue": "/var/lib/postgresql/data/pgdata" }, "PGHOST": { "isOptional": false, "description": "Private-network hostname of the Postgres service.", "defaultValue": "${{RAILWAY_PRIVATE_DOMAIN}}" }, "PGPORT": { "isOptional": false, "description": "Port Postgres listens on.", "defaultValue": "5432" }, "PGUSER": { "isOptional": false, "description": "Postgres superuser name.", "defaultValue": "${{POSTGRES_USER}}" }, "PGDATABASE": { "isOptional": false, "description": "Default database name used by psql and the Postgres client tools.", "defaultValue": "${{POSTGRES_DB}}" }, "PGPASSWORD": { "isOptional": false, "description": "Password for the Postgres superuser.", "defaultValue": "${{POSTGRES_PASSWORD}}" }, "POSTGRES_DB": { "isOptional": false, "description": "Database created on first boot.", "defaultValue": "railway" }, "DATABASE_URL": { "isOptional": false, "description": "Private-network connection string. LiteLLM stores models, virtual keys, budgets and spend history here.", "defaultValue": "postgresql://${{POSTGRES_USER}}:${{POSTGRES_PASSWORD}}@${{RAILWAY_PRIVATE_DOMAIN}}:5432/${{POSTGRES_DB}}" }, "POSTGRES_USER": { "isOptional": false, "description": "Superuser created on first boot.", "defaultValue": "postgres" }, "SSL_CERT_DAYS": { "isOptional": false, "description": "Validity window, in days, for the self-signed TLS certificate this image generates.", "defaultValue": "820" }, "POSTGRES_PASSWORD": { "isOptional": false, "description": "Superuser password, generated once at deploy time.", "defaultValue": "{{POSTGRES_PASSWORD}}" }, "RAILWAY_DEPLOYMENT_DRAINING_SECONDS": { "description": "Seconds Railway keeps the old container alive after the new one is live, so in-flight requests can finish.", "defaultValue": "60" } }, "volumeMounts": { "6d63359f-deb4-46e7-a1c8-08de18538182": { "mountPath": "/var/lib/postgresql/data" } }, "haTemplateCode": "postgres-ha", "haConversionConfig": { "edge": { "label": "Reverse Proxy", "options": [ 2, 3, 4, 5 ], "nodeLabel": "HAProxy", "description": "Routes traffic to the primary", "defaultValue": 3 }, "replica": { "label": "Replicas", "options": [ 2, 3, 4, 5, 6, 7 ], "nodeLabel": "Postgres", "description": "Streaming replicas for automatic failover", "defaultValue": 2 }, "internal": { "label": "Coordinator Nodes", "options": [ 3, 5, 7, 9 ], "nodeLabel": "etcd", "defaultValue": 3 }, "description": "This will convert your Postgres database to an HA cluster with Patroni failover and etcd consensus." } }, "d71b525d-b229-4736-8464-c4cecb79eeb6": { "icon": "https://devicons.railway.app/i/redis.svg", "name": "Redis", "deploy": { "startCommand": "/bin/sh -c 'exec docker-entrypoint.sh redis-server --requirepass \"${REDIS_PASSWORD:?REDIS_PASSWORD must be nonblank}\" --appendonly no --save \"\"'", "healthcheckPath": null }, "source": { "image": "redis:8.2@sha256:7d1e4ce8b9395088377ab382d1f6cfdbd13b3690795198a0399ab8d683064d6d" }, "variables": { "REDISHOST": { "isOptional": false, "description": "Private-network hostname of the Redis service.", "defaultValue": "${{RAILWAY_PRIVATE_DOMAIN}}" }, "REDISPORT": { "isOptional": false, "description": "Port Redis listens on.", "defaultValue": "6379" }, "REDISUSER": { "isOptional": false, "description": "Redis username for this image's default user.", "defaultValue": "default" }, "REDIS_URL": { "isOptional": false, "description": "Private-network connection string. LiteLLM uses Redis for response caching and cross-replica rate limiting.", "defaultValue": "redis://default:${{REDIS_PASSWORD}}@${{RAILWAY_PRIVATE_DOMAIN}}:6379" }, "REDIS_PASSWORD": { "isOptional": false, "description": "Redis password, generated once at deploy time.", "defaultValue": "{{REDIS_PASSWORD}}" }, "RAILWAY_DEPLOYMENT_DRAINING_SECONDS": { "description": "Seconds Railway keeps the old container alive after the new one is live, so in-flight requests can finish.", "defaultValue": "60" } }, "volumeMounts": { "d71b525d-b229-4736-8464-c4cecb79eeb6": { "mountPath": "/data" } } } } } } } } } }, "post_deploy": { "healthcheck": { "service": "LiteLLM", "method": "GET", "path": "/health/readiness", "expect_status": 200 } }, "resources": { "expected_services": 3, "needs_volume": true }, "generated_at": "2026-09-21T10:14:42.516Z", "generator_version": "0.1.0", "status": "validated", "validated_at": "2026-09-21T09:52:50.308Z", "success_rate_30d": 1, "validation": { "last_run_id": "run_8de14ca6a8194e4fb968", "checks": [ { "name": "workflow_completed", "passed": true }, { "name": "all_services_deployed", "passed": true }, { "name": "healthcheck", "passed": true }, { "name": "stays_up", "passed": true } ], "typical_ready_seconds": 108, "typical_build_seconds": 0, "typical_start_seconds": 31, "slowest_service": "LiteLLM" } }