catinspace-au · Nov 15, 2026 · Nov 15, 2026
diff --git a/config.dev.yaml b/config.dev.yaml
@@ -6,7 +6,7 @@
 
 kafka:
   brokers:
-    - localhost:9092
+    - localhost:19092
   group: dfe-loader-dev
   topics:
     - events

diff --git a/docker-compose.dev.yaml b/docker-compose.dev.yaml
@@ -24,7 +24,6 @@ services:
       CLICKHOUSE_PASSWORD: ""
     volumes:
       - clickhouse-data:/var/lib/clickhouse
-      - ./tests/fixtures/init.sql:/docker-entrypoint-initdb.d/init.sql:ro
     healthcheck:
       test: ["CMD", "clickhouse-client", "--query", "SELECT 1"]
       interval: 5s

diff --git a/docs/config-schema.json b/docs/config-schema.json
@@ -195,6 +195,14 @@
         "table": "dlq"
       }
     },
+    "memory": {
+      "description": "Memory limits. **Restart required.**",
+      "$ref": "#/$defs/MemoryConfig",
+      "default": {
+        "limit_bytes": 0,
+        "pressure_threshold": 0.8
+      }
+    },
     "routing": {
       "description": "Table and tenant routing. **Hot-reloaded**, `routing.dlq` aside.",
       "$ref": "#/$defs/RoutingConfig",
@@ -233,14 +241,6 @@
         "flush_age_secs": 5
       }
     },
-    "memory": {
-      "description": "Memory limits. **Hot-reloaded.**",
-      "$ref": "#/$defs/MemoryConfig",
-      "default": {
-        "limit_bytes": 0,
-        "pressure_threshold": 0.8
-      }
-    },
     "timestamp_dq": {
       "description": "Timestamp data quality validation. **Hot-reloaded.**",
       "$ref": "#/$defs/TimestampDqConfig",
@@ -813,7 +813,7 @@
       }
     },
     "InsertFormat": {
-      "description": "Insert format for data writes.\n\nControls how rows are encoded and sent to `ClickHouse`. `RowBinary` is the\ndefault — it uses schema reflection to encode `Map<String, Value>` directly\nto binary, so `ClickHouse` skips JSON parsing entirely. This significantly\nreduces CPU load on the `ClickHouse` cluster at scale.\n\n`JsonEachRow` is the fallback — simpler, self-describing, but the `ClickHouse`\nserver pays the cost of parsing every JSON row on ingest.",
+      "description": "Insert format for data writes.\n\nControls how rows are encoded and sent to `ClickHouse`. `RowBinary` is the\ndefault — it uses schema reflection to encode `Map<String, Value>` directly\nto binary, so `ClickHouse` parses only `JSON` columns, which travel as text.\n\n`JsonEachRow` is the fallback — simpler, self-describing, but the `ClickHouse`\nserver pays the cost of parsing every JSON row on ingest.",
       "oneOf": [
         {
           "description": "Schema-reflected `RowBinary` (default).\n\nFetches schema from `system.columns`, encodes values to binary client-side.\n`ClickHouse` receives pre-columnarised data — zero server-side parsing.\nTotal CPU (client + cluster) is significantly lower than `JSONEachRow`.",
@@ -1367,6 +1367,24 @@
         }
       }
     },
+    "MemoryConfig": {
+      "type": "object",
+      "properties": {
+        "limit_bytes": {
+          "description": "Maximum memory for buffers in bytes. 0 = auto-detect (67% of available)",
+          "type": "integer",
+          "format": "uint",
+          "minimum": 0,
+          "default": 0
+        },
+        "pressure_threshold": {
+          "description": "Pressure threshold (0.0-1.0) - pause consumption above this",
+          "type": "number",
+          "format": "double",
+          "default": 0.8
+        }
+      }
+    },
     "RoutingConfig": {
       "type": "object",
       "properties": {
@@ -1585,24 +1603,6 @@
         }
       }
     },
-    "MemoryConfig": {
-      "type": "object",
-      "properties": {
-        "limit_bytes": {
-          "description": "Maximum memory for buffers in bytes. 0 = auto-detect (67% of available)",
-          "type": "integer",
-          "format": "uint",
-          "minimum": 0,
-          "default": 0
-        },
-        "pressure_threshold": {
-          "description": "Pressure threshold (0.0-1.0) - pause consumption above this",
-          "type": "number",
-          "format": "double",
-          "default": 0.8
-        }
-      }
-    },
     "TimestampDqConfig": {
       "type": "object",
       "properties": {

diff --git a/docs/config-schema.yaml b/docs/config-schema.yaml
@@ -190,6 +190,12 @@ properties:
       - dfe_fetcher_dlq
       - dfe_transform_dlq
       table: dlq
+  memory:
+    description: Memory limits. **Restart required.**
+    $ref: "#/$defs/MemoryConfig"
+    default:
+      limit_bytes: 0
+      pressure_threshold: 0.8
   routing:
     description: Table and tenant routing. **Hot-reloaded**, `routing.dlq` aside.
     $ref: "#/$defs/RoutingConfig"
@@ -221,12 +227,6 @@ properties:
       flush_bytes: 67108864
       flush_rows: 20000
       flush_age_secs: 5
-  memory:
-    description: Memory limits. **Hot-reloaded.**
-    $ref: "#/$defs/MemoryConfig"
-    default:
-      limit_bytes: 0
-      pressure_threshold: 0.8
   timestamp_dq:
     description: Timestamp data quality validation. **Hot-reloaded.**
     $ref: "#/$defs/TimestampDqConfig"
@@ -727,8 +727,7 @@ $defs:
 
       Controls how rows are encoded and sent to `ClickHouse`. `RowBinary` is the
       default — it uses schema reflection to encode `Map<String, Value>` directly
-      to binary, so `ClickHouse` skips JSON parsing entirely. This significantly
-      reduces CPU load on the `ClickHouse` cluster at scale.
+      to binary, so `ClickHouse` parses only `JSON` columns, which travel as text.
 
       `JsonEachRow` is the fallback — simpler, self-describing, but the `ClickHouse`
       server pays the cost of parsing every JSON row on ingest.
@@ -1238,6 +1237,20 @@ $defs:
           picks for its tenant.
         type: string
         default: dlq
+  MemoryConfig:
+    type: object
+    properties:
+      limit_bytes:
+        description: Maximum memory for buffers in bytes. 0 = auto-detect (67% of available)
+        type: integer
+        format: uint
+        minimum: 0
+        default: 0
+      pressure_threshold:
+        description: Pressure threshold (0.0-1.0) - pause consumption above this
+        type: number
+        format: double
+        default: 0.8
   RoutingConfig:
     type: object
     properties:
@@ -1446,20 +1459,6 @@ $defs:
         format: uint64
         minimum: 0
         default: 5
-  MemoryConfig:
-    type: object
-    properties:
-      limit_bytes:
-        description: Maximum memory for buffers in bytes. 0 = auto-detect (67% of available)
-        type: integer
-        format: uint
-        minimum: 0
-        default: 0
-      pressure_threshold:
-        description: Pressure threshold (0.0-1.0) - pause consumption above this
-        type: number
-        format: double
-        default: 0.8
   TimestampDqConfig:
     type: object
     properties:

diff --git a/reference/common_header.csv b/reference/common_header.csv
diff --git a/reference/common_header.sql b/reference/common_header.sql
diff --git a/src/clickhouse/config.rs b/src/clickhouse/config.rs
@@ -52,8 +52,7 @@ impl std::fmt::Display for Transport {
 ///
 /// Controls how rows are encoded and sent to `ClickHouse`. `RowBinary` is the
 /// default — it uses schema reflection to encode `Map<String, Value>` directly
-/// to binary, so `ClickHouse` skips JSON parsing entirely. This significantly
-/// reduces CPU load on the `ClickHouse` cluster at scale.
+/// to binary, so `ClickHouse` parses only `JSON` columns, which travel as text.
 ///
 /// `JsonEachRow` is the fallback — simpler, self-describing, but the `ClickHouse`
 /// server pays the cost of parsing every JSON row on ingest.
@@ -135,7 +134,7 @@ pub struct ClickHouseConfig {
     /// Insert format — how rows are encoded for INSERT.
     ///
     /// `RowBinary` (default): schema-reflected binary encoding. `ClickHouse`
-    /// skips JSON parsing. Lower total CPU across client + cluster.
+    /// parses only `JSON` columns, which travel as text.
     ///
     /// `JsonEachRow`: JSON text via HTTP. Self-describing, `ClickHouse` coerces
     /// types server-side. Higher cluster CPU but simpler.
@@ -352,7 +351,7 @@ impl ClickHouseConfig {
         // JSONEachRow is HTTP-only: the dynamic insert path sends JSONEachRow
         // through `Client::insert_formatted_with` (HTTP). A native-transport
         // client has no HTTP endpoint, so this combination fails at insert
-        // time. Reject it at config time so `config-check` catches it.
+        // time. Rejected here, when the pipeline starts and before readiness.
         if self.transport == Transport::Native && self.insert_format == InsertFormat::JsonEachRow {
             return Err(
                 "insert_format 'json_each_row' requires protocol 'http' (JSONEachRow is \

diff --git a/src/clickhouse/inserter.rs b/src/clickhouse/inserter.rs
@@ -408,7 +408,7 @@ impl Inserter {
     /// Set the insert format.
     ///
     /// - `RowBinary` (default): schema-reflected binary via `DynamicInsert`.
-    ///   `ClickHouse` skips JSON parsing — lower cluster CPU at scale.
+    ///   `ClickHouse` parses only `JSON` columns, which travel as text.
     /// - `JsonEachRow`: NDJSON via `InsertFormatted`. Self-describing fallback.
     #[must_use]
     pub fn with_insert_format(mut self, format: InsertFormat) -> Self {

diff --git a/src/config/loader.rs b/src/config/loader.rs
@@ -89,14 +89,14 @@ pub struct Config {
     pub batch_processing: BatchProcessingConfig,
     /// Kafka DLQ topics loaded into a table of their own. **Restart required.**
     pub dlq_load: DlqLoadConfig,
+    /// Memory limits. **Restart required.**
+    pub memory: MemoryConfig,
 
     // --- Hot-reloaded (takes effect on next batch) ---
     /// Table and tenant routing. **Hot-reloaded**, `routing.dlq` aside.
     pub routing: RoutingConfig,
     /// Buffer flush thresholds. **Hot-reloaded.**
     pub buffer: BufferConfig,
-    /// Memory limits. **Hot-reloaded.**
-    pub memory: MemoryConfig,
     /// Timestamp data quality validation. **Hot-reloaded.**
     pub timestamp_dq: TimestampDqConfig,
     /// Field name sanitisation. **Hot-reloaded.**

diff --git a/src/main.rs b/src/main.rs
@@ -73,8 +73,8 @@ impl ServiceApp for App {
     }
 
     fn load_config(&self, path: Option<&str>) -> Result<Self::Config, CliError> {
-        // run_app has set the cascade up from --config, which also reads
-        // ./.env into the environment Config::load reads.
+        // run_app has set the cascade up from --config. Without --config it
+        // also reads ./.env into the environment Config::load reads.
         let config = Config::load(path).map_err(|e| CliError::Config(e.to_string()))?;
         config
             .validate()
@@ -128,9 +128,8 @@ impl ServiceApp for App {
                 .clone()
                 .unwrap_or_else(|| Arc::new(config.scaling.build_pressure()));
 
-            // The engine KEDA reads takes its gates from scalo's cascade, which
-            // reads neither the `--config` file nor the `DFE_LOADER__` env form,
-            // so a gate set either way is reported here and ignored there (#160).
+            // The engine KEDA reads takes its gates from scalo's cascade, so a
+            // gate it did not pick up is reported here.
             if scaling.is_enabled() != config.scaling.enabled {
                 warn!(
                     configured = config.scaling.enabled,