Improve jupyter startup time (#53)

* Remove jupyterlab to improve startup time

* Disable unused jupyter extensions

* Increase sandbox timeout to 20 minutes

* Add changeset

* Use tighter exponential backoff

* Delay pre-warming requests

Author: ghostwriternr

.changeset/thirty-months-report.md

@@ -0,0 +1,5 @@
+---
+"@cloudflare/sandbox": patch
+---
+
+Improve jupyterlab config to speed up startup

packages/sandbox/Dockerfile

@@ -37,6 +37,7 @@ RUN apt-get update && apt-get install -y \
     ca-certificates \
     gnupg \
     lsb-release \
+    strace \
     && rm -rf /var/lib/apt/lists/*
 
 # Set Python 3.11 as default python3
@@ -56,17 +57,20 @@ RUN apt-get update && apt-get install -y ca-certificates curl gnupg \
 COPY --from=bun-source /usr/local/bin/bun /usr/local/bin/bun
 COPY --from=bun-source /usr/local/bin/bunx /usr/local/bin/bunx
 
-# Install Jupyter and kernels
+# Install minimal Jupyter components
 RUN pip3 install --no-cache-dir \
-    jupyter \
-    jupyterlab \
+    jupyter-server \
+    jupyter-client \
     ipykernel \
-    notebook \
+    orjson \
+    && python3 -m ipykernel install --user --name python3
+
+# Install scientific packages
+RUN pip3 install --no-cache-dir \
     matplotlib \
     numpy \
     pandas \
-    seaborn \
-    && python3 -m ipykernel install --user --name python3
+    seaborn
 
 # Install JavaScript kernel (ijavascript) - using E2B's fork
 RUN npm install -g --unsafe-perm git+https://github.com/e2b-dev/ijavascript.git \

packages/sandbox/container_src/jupyter-service.ts

@@ -133,16 +133,26 @@ export class JupyterService {
    * Pre-warm context pools for better performance
    */
   private async warmContextPools() {
+    // Delay pre-warming to avoid startup rush that triggers Bun WebSocket bug
+    await new Promise((resolve) => setTimeout(resolve, 2000));
+
     try {
       console.log(
         "[JupyterService] Pre-warming context pools for better performance"
       );
 
-      // Enable pool warming with min sizes
+      // Warm one at a time with delay between them
       await this.jupyterServer.enablePoolWarming("python", 1);
+
+      // Small delay between different language kernels
+      await new Promise((resolve) => setTimeout(resolve, 1000));
+
       await this.jupyterServer.enablePoolWarming("javascript", 1);
     } catch (error) {
       console.error("[JupyterService] Error pre-warming context pools:", error);
+      console.error(
+        "[JupyterService] Pre-warming failed but service continues"
+      );
     }
   }
 

packages/sandbox/container_src/jupyter_config.py

@@ -0,0 +1,48 @@
+"""
+Minimal Jupyter configuration focused on kernel-only usage
+"""
+
+c = get_config()  # noqa
+
+# Disable all authentication - we handle security at container level
+c.ServerApp.token = ''
+c.ServerApp.password = ''
+c.IdentityProvider.token = ''
+c.ServerApp.allow_origin = '*'
+c.ServerApp.allow_remote_access = True
+c.ServerApp.disable_check_xsrf = True
+c.ServerApp.allow_root = True
+c.ServerApp.allow_credentials = True
+
+# Also set NotebookApp settings for compatibility
+c.NotebookApp.token = ''
+c.NotebookApp.password = ''
+c.NotebookApp.allow_origin = '*'
+c.NotebookApp.allow_remote_access = True
+c.NotebookApp.disable_check_xsrf = True
+c.NotebookApp.allow_credentials = True
+
+# Performance settings
+c.ServerApp.iopub_data_rate_limit = 1000000000  # E2B uses 1GB/s
+
+# Minimal logging
+c.Application.log_level = 'ERROR'
+
+# Disable browser
+c.ServerApp.open_browser = False
+
+# Optimize for container environment
+c.ServerApp.ip = '0.0.0.0'
+c.ServerApp.port = 8888
+
+# Kernel optimizations
+c.KernelManager.shutdown_wait_time = 0.0
+c.MappingKernelManager.cull_idle_timeout = 0
+c.MappingKernelManager.cull_interval = 0
+
+# Disable terminals
+c.ServerApp.terminals_enabled = False
+
+# Disable all extensions to speed up startup
+c.ServerApp.jpserver_extensions = {}
+c.ServerApp.nbserver_extensions = {}

packages/sandbox/container_src/startup.sh

@@ -2,7 +2,8 @@
 
 # Function to check if Jupyter is ready
 check_jupyter_ready() {
-  curl -s http://localhost:8888/api > /dev/null 2>&1
+  # Check if API is responsive and kernelspecs are available
+  curl -s http://localhost:8888/api/kernelspecs > /dev/null 2>&1
 }
 
 # Function to notify Bun server that Jupyter is ready
@@ -12,19 +13,10 @@ notify_jupyter_ready() {
   echo "[Startup] Jupyter is ready, notified Bun server"
 }
 
-# Start Jupyter notebook server in background
+# Start Jupyter server in background
 echo "[Startup] Starting Jupyter server..."
-jupyter notebook \
-  --ip=0.0.0.0 \
-  --port=8888 \
-  --no-browser \
-  --allow-root \
-  --NotebookApp.token='' \
-  --NotebookApp.password='' \
-  --NotebookApp.allow_origin='*' \
-  --NotebookApp.disable_check_xsrf=True \
-  --NotebookApp.allow_remote_access=True \
-  --NotebookApp.allow_credentials=True \
+jupyter server \
+  --config=/container-server/jupyter_config.py \
   > /tmp/jupyter.log 2>&1 &
 
 JUPYTER_PID=$!
@@ -37,40 +29,49 @@ BUN_PID=$!
 # Monitor Jupyter readiness in background
 (
   echo "[Startup] Monitoring Jupyter readiness in background..."
-  MAX_ATTEMPTS=30
+  MAX_ATTEMPTS=60
   ATTEMPT=0
-  DELAY=0.5
-  MAX_DELAY=5
-  
+
+  # Track start time for reporting
+  START_TIME=$(date +%s.%N)
+
   while [ $ATTEMPT -lt $MAX_ATTEMPTS ]; do
     if check_jupyter_ready; then
       notify_jupyter_ready
-      echo "[Startup] Jupyter server is ready after $ATTEMPT attempts"
+      END_TIME=$(date +%s.%N)
+      ELAPSED=$(awk "BEGIN {printf \"%.2f\", $END_TIME - $START_TIME}")
+      echo "[Startup] Jupyter server is ready after $ELAPSED seconds ($ATTEMPT attempts)"
       break
     fi
-    
+
     # Check if Jupyter process is still running
     if ! kill -0 $JUPYTER_PID 2>/dev/null; then
       echo "[Startup] WARNING: Jupyter process died. Check /tmp/jupyter.log for details"
       cat /tmp/jupyter.log
       # Don't exit - let Bun server continue running in degraded mode
       break
     fi
-    
+
     ATTEMPT=$((ATTEMPT + 1))
-    echo "[Startup] Jupyter not ready yet (attempt $ATTEMPT/$MAX_ATTEMPTS, delay ${DELAY}s)"
-    
-    # Sleep with exponential backoff
-    sleep $DELAY
-    
-    # Increase delay exponentially with jitter, cap at MAX_DELAY
-    DELAY=$(awk "BEGIN {printf \"%.2f\", $DELAY * 1.5 + (rand() * 0.5)}")
-    # Use awk for comparison since bc might not be available
-    if [ $(awk "BEGIN {print ($DELAY > $MAX_DELAY)}") -eq 1 ]; then
-      DELAY=$MAX_DELAY
+
+    # Start with faster checks
+    if [ $ATTEMPT -eq 1 ]; then
+      DELAY=0.5  # Start at 0.5s
+    else
+      # Exponential backoff with 1.3x multiplier (less aggressive than 1.5x)
+      DELAY=$(awk "BEGIN {printf \"%.2f\", $DELAY * 1.3}")
+      # Cap at 2s max (instead of 5s)
+      if [ $(awk "BEGIN {print ($DELAY > 2)}") -eq 1 ]; then
+        DELAY=2
+      fi
     fi
+
+    # Log with current delay for transparency
+    echo "[Startup] Jupyter not ready yet (attempt $ATTEMPT/$MAX_ATTEMPTS, next check in ${DELAY}s)"
+
+    sleep $DELAY
   done
-  
+
   if [ $ATTEMPT -eq $MAX_ATTEMPTS ]; then
     echo "[Startup] WARNING: Jupyter failed to become ready within attempts"
     echo "[Startup] Jupyter logs:"
@@ -80,4 +81,4 @@ BUN_PID=$!
 ) &
 
 # Wait for Bun server (main process)
-wait $BUN_PID
+wait $BUN_PID

packages/sandbox/src/sandbox.ts

@@ -37,7 +37,7 @@ export function getSandbox(ns: DurableObjectNamespace<Sandbox>, id: string) {
 
 export class Sandbox<Env = unknown> extends Container<Env> implements ISandbox {
   defaultPort = 3000; // Default port for the container's Bun server
-  sleepAfter = "3m"; // Sleep the sandbox if no requests are made in this timeframe
+  sleepAfter = "20m"; // Keep container warm for 20 minutes to avoid cold starts
   client: JupyterClient;
   private sandboxName: string | null = null;
   private codeInterpreter: CodeInterpreter;