fix(executor): mark pool backend experimental, auto selects ephemeral

docs/self-hosting/security.mdx

@@ -59,7 +59,7 @@ To enable nsjail, set the following in your `.env`:
 
 ```bash
 TRACECAT__DISABLE_NSJAIL=false
-TRACECAT__EXECUTOR_BACKEND=pool  # or 'ephemeral' for multi-tenant full isolation
+TRACECAT__EXECUTOR_BACKEND=ephemeral
 ```
 
 nsjail requires:
@@ -76,8 +76,15 @@ nsjail requires:
 | Backend | Isolation | Latency | Use case |
 | :------ | :-------- | :------ | :------- |
 | `direct` | PID namespace only | ~50ms | Development, trusted environments |
-| `pool` | nsjail sandbox (warm workers) | ~100-200ms | Single-tenant production |
-| `ephemeral` | nsjail sandbox (cold per action) | ~4000ms | Multi-tenant, maximum isolation |
+| `ephemeral` | nsjail sandbox (cold per action) | ~4000ms | Production, maximum isolation |
+| `pool` | nsjail sandbox (warm workers) | ~100-200ms | Experimental — not production ready |
+
+<Warning>
+  The `pool` backend is experimental and not production ready. Its warm workers
+  import from registry cache paths that eviction cannot reclaim, so the
+  executor's registry cache can grow without bound. Use `ephemeral` for
+  production deployments.
+</Warning>
 
 ## Authentication
 

docs/snippets/environment-variables.mdx

@@ -69,7 +69,7 @@
 
 | Variable | Default | Description |
 | :------- | :------ | :---------- |
-| `TRACECAT__EXECUTOR_BACKEND` | `direct` | Execution strategy for actions. `direct` runs a subprocess per action, `pool` uses warm nsjail workers (requires nsjail), `ephemeral` spawns a cold nsjail subprocess per action for full isolation, `auto` selects `pool` if nsjail is available and falls back to `direct`, `test` runs in-process for tests only. |
+| `TRACECAT__EXECUTOR_BACKEND` | `direct` | Execution strategy for actions. `direct` runs a subprocess per action, `ephemeral` spawns a cold nsjail subprocess per action for full isolation, `auto` selects `ephemeral` if nsjail is available and falls back to `direct`, `pool` uses warm nsjail workers (experimental, not production ready), `test` runs in-process for tests only. |
 | `TRACECAT__DISABLE_NSJAIL` | `true` | Disable nsjail sandboxing. Set to `false` only if nsjail is installed. |
 | `TRACECAT__RESULT_EXTERNALIZATION_ENABLED` | `true` | Store large action results in blob storage instead of Temporal history. |
 | `TRACECAT__COLLECTION_MANIFESTS_ENABLED` | `true` | Store large collections as chunked manifests in blob storage. |

tracecat/config.py

@@ -636,11 +636,13 @@ def _parse_auth_types() -> set[AuthType]:
 """Executor backend for running actions.
 
 Supported values:
-- 'pool': Warm nsjail workers (single-tenant, high throughput, ~100-200ms)
+- 'pool': Warm nsjail workers (single-tenant, high throughput, ~100-200ms).
+  EXPERIMENTAL: not production ready. The registry cache is exempt from
+  eviction under this backend and can grow without bound.
 - 'ephemeral': Cold nsjail subprocess per action (multitenant, full isolation, ~4000ms)
 - 'direct': Direct subprocess execution (no warm workers, no in-process state sharing)
 - 'test': In-process execution for tests only (no isolation, no subprocess overhead)
-- 'auto': Auto-select based on environment (pool if nsjail available, else direct)
+- 'auto': Auto-select based on environment (ephemeral if nsjail available, else direct)
 
 Trust mode is derived from the backend type:
 - pool: untrusted (secrets pre-resolved, no DB creds)

tracecat/executor/backends/pool/backend.py

@@ -1,5 +1,10 @@
 """Worker pool executor backend.
 
+EXPERIMENTAL: not production ready. Warm workers import from registry cache
+paths that in-process leases cannot see, so tarball cache entries are exempt
+from eviction and the cache can grow without bound. Use the ephemeral backend
+for production nsjail isolation.
+
 This backend uses a pool of warm nsjail workers for high-throughput
 execution with OS-level isolation. Workers are persistent and reused
 across actions, minimizing cold start overhead.
@@ -32,6 +37,9 @@
 class PoolBackend(ExecutorBackend):
     """Warm nsjail worker pool backend.
 
+    EXPERIMENTAL: not production ready. See the module docstring for the
+    registry cache eviction limitation.
+
     Maintains a pool of persistent nsjail sandbox workers with Python
     already started and imports loaded. This provides:
 

tracecat/executor/schemas.py

@@ -40,11 +40,14 @@ class ExecutorBackendType(StrEnum):
 
     All sandbox backends use untrusted mode - DB credentials are never passed.
 
-    - POOL: Warm nsjail workers (high throughput, single-tenant, untrusted)
+    - POOL: Warm nsjail workers (high throughput, single-tenant, untrusted).
+      EXPERIMENTAL: not production ready. Warm workers import from registry
+      cache paths that in-process leases cannot see, so tarball cache entries
+      are exempt from eviction and the cache can grow without bound.
     - EPHEMERAL: Cold nsjail subprocess per action (full isolation, multi-tenant, untrusted)
     - DIRECT: Direct subprocess execution (no warm workers)
     - TEST: In-process execution for tests only
-    - AUTO: Auto-select based on environment
+    - AUTO: Auto-select based on environment (never selects experimental backends)
     """
 
     POOL = "pool"
@@ -85,14 +88,20 @@ def resolve_backend_type() -> ExecutorBackendType:
             backend_type = ExecutorBackendType.DIRECT
         elif _is_nsjail_available():
             logger.info(
-                "Auto-selecting 'pool' backend (nsjail available)",
+                "Auto-selecting 'ephemeral' backend (nsjail available)",
             )
-            backend_type = ExecutorBackendType.POOL
+            backend_type = ExecutorBackendType.EPHEMERAL
         else:
             logger.warning(
                 "Auto-selecting 'direct' backend (nsjail not available)",
             )
             backend_type = ExecutorBackendType.DIRECT
+    elif backend_type == ExecutorBackendType.POOL:
+        logger.warning(
+            "The 'pool' executor backend is experimental and not production "
+            "ready: its registry cache is exempt from eviction and can grow "
+            "without bound. Use 'ephemeral' for production nsjail isolation.",
+        )
 
     return backend_type