Update to MessagePack columnar format and latest performance numbers

- Updated write performance from 2.01M to 2.32M RPS
- Changed MessagePack examples to columnar format (9.7x faster)
- Added batch example for multiple rows
- Updated write benchmarks with columnar vs row format comparison
- Added latency improvements (20x lower p50, 21x lower p95, 26x lower p99)

Generated with Claude Code

Co-Authored-By: Claude <noreply@anthropic.com>

Author: Ignacio Van Droogenbroeck

docs/getting-started.md

@@ -16,7 +16,7 @@ This guide will get you up and running with Arc in 5 minutes.
 
 ### Option 1: Native Deployment (Recommended)
 
-Native deployment provides **3.5x faster performance** than Docker (2.01M RPS vs 570K RPS).
+Native deployment provides **4x faster performance** than Docker (2.32M RPS vs 570K RPS).
 
 ```bash
 # Clone the repository
@@ -77,7 +77,9 @@ export ARC_TOKEN="your-token-here"
 
 ## Write Your First Data
 
-### Using MessagePack (Recommended - 7.9x Faster)
+### Using MessagePack Columnar (Recommended - 9.7x Faster)
+
+MessagePack columnar format provides the best performance (2.32M RPS vs 240K Line Protocol).
 
 ```python
 import msgpack
@@ -87,24 +89,18 @@ import os
 
 token = os.getenv("ARC_TOKEN")
 
-# Prepare data
+# Columnar format - arrange data by columns (fastest)
 data = {
-    "batch": [
-        {
-            "m": "cpu",                                      # measurement name
-            "t": int(datetime.now().timestamp() * 1000),    # timestamp (milliseconds)
-            "h": "server01",                                 # host tag
-            "tags": {                                        # additional tags
-                "region": "us-east",
-                "dc": "aws"
-            },
-            "fields": {                                      # metric values
-                "usage_idle": 95.0,
-                "usage_user": 3.2,
-                "usage_system": 1.8
-            }
-        }
-    ]
+    "m": "cpu",                                      # measurement name
+    "columns": {
+        "time": [int(datetime.now().timestamp() * 1000)],   # timestamps
+        "host": ["server01"],                                # host tag
+        "region": ["us-east"],                               # region tag
+        "dc": ["aws"],                                       # dc tag
+        "usage_idle": [95.0],                                # metric value
+        "usage_user": [3.2],                                 # metric value
+        "usage_system": [1.8]                                # metric value
+    }
 }
 
 # Send data
@@ -123,6 +119,35 @@ else:
     print(f"Error {response.status_code}: {response.text}")
 ```
 
+**Batch multiple rows for even better performance:**
+
+```python
+# Send multiple rows in one request
+data = {
+    "m": "cpu",
+    "columns": {
+        "time": [
+            int(datetime.now().timestamp() * 1000),
+            int(datetime.now().timestamp() * 1000),
+            int(datetime.now().timestamp() * 1000)
+        ],
+        "host": ["server01", "server02", "server03"],
+        "usage_idle": [95.0, 87.5, 92.3],
+        "usage_user": [3.2, 8.1, 5.4],
+        "usage_system": [1.8, 4.4, 2.3]
+    }
+}
+
+response = requests.post(
+    "http://localhost:8000/write/v2/msgpack",
+    headers={
+        "Authorization": f"Bearer {token}",
+        "Content-Type": "application/msgpack"
+    },
+    data=msgpack.packb(data)
+)
+```
+
 ### Using InfluxDB Line Protocol
 
 ```bash

docs/performance/benchmarks.md

@@ -187,23 +187,31 @@ Arc excels at:
 
 ## Write Performance
 
-Arc achieves exceptional write throughput through MessagePack binary protocol.
+Arc achieves exceptional write throughput through MessagePack columnar binary protocol.
 
-### Write Benchmarks
+### Write Benchmarks - Format Comparison
 
-| Storage Backend | Throughput | p50 Latency | p95 Latency | p99 Latency |
-|----------------|------------|-------------|-------------|-------------|
-| **Local NVMe** | **2.08M RPS** | **13.4ms** | **136ms** | **280ms** |
-| **MinIO** | **2.01M RPS** | **16.6ms** | **147ms** | **318ms** |
-| **Line Protocol** | **240K RPS** | N/A | N/A | N/A |
+| Wire Format | Throughput | p50 Latency | p95 Latency | p99 Latency | Notes |
+|-------------|------------|-------------|-------------|-------------|-------|
+| **MessagePack Columnar** | **2.32M RPS** | **6.75ms** | **39.46ms** | **59.09ms** | Zero-copy passthrough (RECOMMENDED) |
+| **MessagePack Row** | **908K RPS** | **136.86ms** | **851.71ms** | **1542ms** | Legacy format with conversion overhead |
+| **Line Protocol** | **240K RPS** | N/A | N/A | N/A | InfluxDB compatibility mode |
+
+**Columnar Format Advantages:**
+- **2.55x faster throughput** vs row format (2.32M vs 908K RPS)
+- **20x lower p50 latency** (6.75ms vs 136.86ms)
+- **21x lower p95 latency** (39.46ms vs 851.71ms)
+- **26x lower p99 latency** (59.09ms vs 1542ms)
+- **67x fewer errors** under load (63 vs 4,211 errors at 2.5M target RPS)
 
 **Test Configuration**:
 - Hardware: Apple M3 Max (14 cores)
-- Workers: 42 (3x CPU cores)
-- Protocol: MessagePack binary streaming
+- Workers: 400
+- Protocol: MessagePack columnar binary streaming
 - Deployment: Native mode
+- Storage: MinIO
 
-**MessagePack vs Line Protocol**: 8.4x faster
+**MessagePack Columnar vs Line Protocol**: 9.7x faster
 
 ## Query Format Performance