confluent-kafka-tornado.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Example using confluent-kafka-python with Tornado\n",
    "\n",
    "Tornado is a common async framework in Python.  It enables concurrency among many non-blocking applications.\n",
    "\n",
    "Confluent-kafka-python is a Kafka framework by Confluent, the main for-profit entity in this space.  It is known to be fast and follow the Java API fairly consistently.  It is normally a blocking application.\n",
    "\n",
    "This notebook shows two approaches to provide non-blocking tornado-friendly operation with the `confluent_kafka` library."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tornado import gen\n",
    "from tornado.ioloop import IOLoop\n",
    "import confluent_kafka as ck"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Connect to topic 'foo', set all partitions to offset 0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "consumer = ck.Consumer({'bootstrap.servers': 'localhost:9092', 'group.id': 'mygroup-foo2'})\n",
    "parts = [ck.TopicPartition('foo', i, 0) for i in range(4)]\n",
    "consumer.assign(parts)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Time baseline performance with blocking API"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 160 ms, sys: 0 ns, total: 160 ms\n",
      "Wall time: 162 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "L = []\n",
    "\n",
    "for i in range(100000):\n",
    "    msg = consumer.poll(2)\n",
    "    L.append(msg.value())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### First attempt with Tornado\n",
    "\n",
    "We build a Tornado coroutine to get data from our consumer in a non-blocking fashion.  Our approach is as follows:\n",
    "\n",
    "1.  If data is available immediately then return it\n",
    "2.  If data is not available then call the blocking `poll` function in a separate thread\n",
    "\n",
    "Note that consumers are not threadsafe.  We handle this on the tornado side by only having one thread so concurrent calls to this function will process in a first-come-first-served fasion.   We assume (perhaps erroneously) that the user isn't calling `consumer.poll` on their own in other threads."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "from concurrent.futures import ThreadPoolExecutor\n",
    "from time import time\n",
    "\n",
    "e = ThreadPoolExecutor(1)  # Use a single thread"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tornado import gen\n",
    "\n",
    "@gen.coroutine\n",
    "def poll(consumer, timeout=1e9):\n",
    "    msg = consumer.poll(0)\n",
    "    if msg is None:\n",
    "        msg = yield e.submit(consumer.poll, timeout)\n",
    "    return msg"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Profile Tornado first attempt\n",
    "\n",
    "We reset our topics and make a tornado callback that uses our `poll` function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.124321699142456 seconds\n"
     ]
    }
   ],
   "source": [
    "parts = [ck.TopicPartition('foo', i, 0) for i in range(4)]\n",
    "consumer.assign(parts)\n",
    "\n",
    "L = []\n",
    "\n",
    "@gen.coroutine\n",
    "def poll_many(timeout=1e9):\n",
    "    start = time()\n",
    "    for i in range(100000):\n",
    "        msg = yield poll(consumer, timeout=timeout)\n",
    "        L.append(msg.value())\n",
    "    end = time()\n",
    "    print(end - start, 'seconds')\n",
    "    \n",
    "IOLoop.current().add_callback(poll_many, 10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This runs a bit slower than the pure `confluent_kafka` solution.  Tornado's event loop does introduce some overhead.  In *many* cases, this is fine.  We're still operating at around 70,000 messages per second, which is more than adequate for many applications."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Tornado: second attempt with batching\n",
    "\n",
    "We can reduce this overhead by batching.  Every call to `poll_batch` now returns a list of messages.  This list has at least one element and no more than `batchsize` elements.  We use the same trick with the thread pool executor."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "@gen.coroutine\n",
    "def poll_batch(consumer, timeout=1e9, batchsize=100):\n",
    "    msg = consumer.poll(0)\n",
    "    if msg is None:\n",
    "        msg = yield e.submit(consumer.poll, timeout)\n",
    "    batch = [msg]\n",
    "    while len(batch) < 100:\n",
    "        msg = consumer.poll(0)\n",
    "        if msg is None:\n",
    "            break\n",
    "        else:\n",
    "            batch.append(msg)\n",
    "    return batch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.3510932922363281 seconds\n"
     ]
    }
   ],
   "source": [
    "# Reset offsets to zero\n",
    "parts = [ck.TopicPartition('foo', i, 0) for i in range(4)]\n",
    "consumer.assign(parts)\n",
    "\n",
    "L = []\n",
    "\n",
    "@gen.coroutine\n",
    "def poll_many(timeout=1e9):\n",
    "    start = time()\n",
    "    while len(L) < 100000:\n",
    "        batch = yield poll_batch(consumer, timeout=timeout, batchsize=100)\n",
    "        L.extend([msg.value() for msg in batch])\n",
    "    end = time()\n",
    "    print(end - start, 'seconds')\n",
    "\n",
    "IOLoop.current().add_callback(poll_many, 10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This is faster.  I see times that get all the way down to normal blocking speed. I also see times that are closer to a second.  I attribute this variation to other things running on the event loop, but I'm not sure.\n",
    "\n",
    "This *can* get to full speed."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Other options\n",
    "\n",
    "Rather than rely on an external thread, we could register a `tornado.locks.Event` object that was triggered every time the internal buffer for a partition went from the empty to non-empty state.  This would require changes within the `confluent_kafka` codebase to allow callbacks at various points during operation."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

producer.ipynb