RSVP for your your local TensorFlow Everywhere event today!

Streaming structured data from Elasticsearch using Tensorflow-IO

View on Run in Google Colab View source on GitHub Download notebook


This tutorial focuses on streaming data from an Elasticsearch cluster into a which is then used in conjunction with tf.keras for training and inference.

Elasticseach is primarily a distributed search engine which supports storing structured, unstructured, geospatial, numeric data etc. For the purpose of this tutorial, a dataset with structured records is utilized.

Setup packages

The elasticsearch package is utilized for preparing and storing the data within elasticsearch indices for demonstration purposes only. In real-world production clusters with numerous nodes, the cluster might be receiving the data from connectors like logstash etc.

Once the data is available in the elasticsearch cluster, only tensorflow-io is required to stream the data into the models.

Install the required tensorflow-io and elasticsearch packages

pip install tensorflow-io
pip install elasticsearch

Import packages

import os
import time
from sklearn.model_selection import train_test_split
from elasticsearch import Elasticsearch
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.layers.experimental import preprocessing
import tensorflow_io as tfio

Validate tf and tfio imports

print("tensorflow-io version: {}".format(tfio.__version__))
print("tensorflow version: {}".format(tf.__version__))
tensorflow-io version: 0.16.0
tensorflow version: 2.3.0

Download and setup the Elasticsearch instance

For demo purposes, the open-source version of the elasticsearch package is used.

wget -q
wget -q
tar -xzf elasticsearch-oss-7.9.2-linux-x86_64.tar.gz
sudo chown -R daemon:daemon elasticsearch-7.9.2/
shasum -a 512 -c elasticsearch-oss-7.9.2-linux-x86_64.tar.gz.sha512
elasticsearch-oss-7.9.2-linux-x86_64.tar.gz: OK

Run the instance as a daemon process

sudo -H -u daemon elasticsearch-7.9.2/bin/elasticsearch
Starting job # 0 in a separate thread.

# Sleep for few seconds to let the instance start.

Once the instance has been started, grep for elasticsearch in the processes list to confirm the availability.

ps -ef | grep elasticsearch
root         144     142  0 21:24 ?        00:00:00 sudo -H -u daemon elasticsearch-7.9.2/bin/elasticsearch
daemon       145     144 86 21:24 ?        00:00:17 /content/elasticsearch-7.9.2/jdk/bin/java -Xshare:auto -Des.networkaddress.cache.ttl=60 -Des.networkaddress.cache.negative.ttl=10 -XX:+AlwaysPreTouch -Xss1m -Djava.awt.headless=true -Dfile.encoding=UTF-8 -Djna.nosys=true -XX:-OmitStackTraceInFastThrow -XX:+ShowCodeDetailsInExceptionMessages -Dio.netty.noUnsafe=true -Dio.netty.noKeySetOptimization=true -Dio.netty.recycler.maxCapacityPerThread=0 -Dio.netty.allocator.numDirectArenas=0 -Dlog4j.shutdownHookEnabled=false -Dlog4j2.disable.jmx=true -Djava.locale.providers=SPI,COMPAT -Xms1g -Xmx1g -XX:+UseG1GC -XX:G1ReservePercent=25 -XX:InitiatingHeapOccupancyPercent=30 -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=data -XX:ErrorFile=logs/hs_err_pid%p.log -Xlog:gc*,gc+age=trace,safepoint:file=logs/gc.log:utctime,pid,tags:filecount=32,filesize=64m -XX:MaxDirectMemorySize=536870912 -Des.path.home=/content/elasticsearch-7.9.2 -Des.path.conf=/content/elasticsearch-7.9.2/config -Des.distribution.flavor=oss -Des.distribution.type=tar -Des.bundled_jdk=true -cp /content/elasticsearch-7.9.2/lib/* org.elasticsearch.bootstrap.Elasticsearch
root         382     380  0 21:24 ?        00:00:00 grep elasticsearch

query the base endpoint to retrieve information about the cluster.

curl -sX GET "localhost:9200/"
  "name" : "d1bc7d054c69",
  "cluster_name" : "elasticsearch",
  "cluster_uuid" : "P8YXfKqYS-OS3k9CdMmlsw",
  "version" : {
    "number" : "7.9.2",
    "build_flavor" : "oss",
    "build_type" : "tar",
    "build_hash" : "d34da0ea4a966c4e49417f2da2f244e3e97b4e6e",
    "build_date" : "2020-09-23T00:45:33.626720Z",
    "build_snapshot" : false,
    "lucene_version" : "8.6.2",
    "minimum_wire_compatibility_version" : "6.8.0",
    "minimum_index_compatibility_version" : "6.0.0-beta1"
  "tagline" : "You Know, for Search"

Explore the dataset

For the purpose of this tutorial, lets download the PetFinder dataset and feed the data into elasticsearch manually. The goal of this classification problem is predict if the pet will be adopted or not.

dataset_url = ''
csv_file = 'datasets/petfinder-mini/petfinder-mini.csv'
tf.keras.utils.get_file('', dataset_url,
                        extract=True, cache_dir='.')
pf_df = pd.read_csv(csv_file)
Downloading data from
1671168/1668792 [==============================] - 0s 0us/step


For the purpose of the tutorial, modifications are made to the label column. 0 will indicate the pet was not adopted, and 1 will indicate that it was.

# In the original dataset "4" indicates the pet was not adopted.
pf_df['target'] = np.where(pf_df['AdoptionSpeed']==4, 0, 1)

# Drop un-used columns.
pf_df = pf_df.drop(columns=['AdoptionSpeed', 'Description'])
# Number of datapoints and columns
len(pf_df), len(pf_df.columns)
(11537, 14)

Split the dataset

train_df, test_df = train_test_split(pf_df, test_size=0.3, shuffle=True)
print("Number of training samples: ",len(train_df))
print("Number of testing sample: ",len(test_df))
Number of training samples:  8075
Number of testing sample:  3462

Store the train and test data in elasticsearch indices

Storing the data in the local elasticsearch cluster simulates an environment for continuous remote data retrieval for training and inference purposes.

ES_NODES = "http://localhost:9200"

def prepare_es_data(index, doc_type, df):
  records = df.to_dict(orient="records")
  es_data = []
  for idx, record in enumerate(records):
    meta_dict = {
          "index": {
              "_index": index, 
              "_type": doc_type, 
              "_id": idx

  return es_data

def index_es_data(index, es_data):
  es_client = Elasticsearch(hosts = [ES_NODES])
  if es_client.indices.exists(index):
      print("deleting the '{}' index.".format(index))
      res = es_client.indices.delete(index=index)
      print("Response from server: {}".format(res))

  print("creating the '{}' index.".format(index))
  res = es_client.indices.create(index=index)
  print("Response from server: {}".format(res))

  print("bulk index the data")
  res = es_client.bulk(index=index, body=es_data, refresh = True)
  print("Errors: {}, Num of records indexed: {}".format(res["errors"], len(res["items"])))
train_es_data = prepare_es_data(index="train", doc_type="pet", df=train_df)
test_es_data = prepare_es_data(index="test", doc_type="pet", df=test_df)

index_es_data(index="train", es_data=train_es_data)
index_es_data(index="test", es_data=test_es_data)
creating the 'train' index.
Response from server: {'acknowledged': True, 'shards_acknowledged': True, 'index': 'train'}
bulk index the data

/usr/local/lib/python3.6/dist-packages/elasticsearch/connection/ ElasticsearchDeprecationWarning: [types removal] Specifying types in bulk requests is deprecated.
  warnings.warn(message, category=ElasticsearchDeprecationWarning)

Errors: False, Num of records indexed: 8075
creating the 'test' index.
Response from server: {'acknowledged': True, 'shards_acknowledged': True, 'index': 'test'}
bulk index the data
Errors: False, Num of records indexed: 3462

Prepare tfio datasets

Once the data is available in the cluster, only tensorflow-io is required to stream the data from the indices. The elasticsearch.ElasticsearchIODataset class is utilized for this purpose. The class inherits from and thus exposes all the useful functionalities of out of the box.

Training dataset

HEADERS = {"Content-Type": "application/json"}

train_ds = tfio.experimental.elasticsearch.ElasticsearchIODataset(

# Prepare a tuple of (features, label)
train_ds = v: (v, v.pop("target")))
train_ds = train_ds.batch(BATCH_SIZE)
Connection successful: http://localhost:9200/_cluster/health

Testing dataset

test_ds = tfio.experimental.elasticsearch.ElasticsearchIODataset(

# Prepare a tuple of (features, label)
test_ds = v: (v, v.pop("target")))
test_ds = test_ds.batch(BATCH_SIZE)
Connection successful: http://localhost:9200/_cluster/health

Define the keras preprocessing layers

As per the structured data tutorial, it is recommended to use the Keras Preprocessing Layers as they are more intuitive, and can be easily integrated with the models. However, the standard feature_columns can also be used.

For a better understanding of the preprocessing_layers in classifying structured data, please refer to the structured data tutorial

def get_normalization_layer(name, dataset):
  # Create a Normalization layer for our feature.
  normalizer = preprocessing.Normalization()

  # Prepare a Dataset that only yields our feature.
  feature_ds = x, y: x[name])

  # Learn the statistics of the data.

  return normalizer

def get_category_encoding_layer(name, dataset, dtype, max_tokens=None):
  # Create a StringLookup layer which will turn strings into integer indices
  if dtype == 'string':
    index = preprocessing.StringLookup(max_tokens=max_tokens)
    index = preprocessing.IntegerLookup(max_values=max_tokens)

  # Prepare a Dataset that only yields our feature
  feature_ds = x, y: x[name])

  # Learn the set of possible values and assign them a fixed integer index.

  # Create a Discretization for our integer indices.
  encoder = preprocessing.CategoryEncoding(max_tokens=index.vocab_size())

  # Prepare a Dataset that only yields our feature.
  feature_ds =

  # Learn the space of possible indices.

  # Apply one-hot encoding to our indices. The lambda function captures the
  # layer so you can use them, or include them in the functional model later.
  return lambda feature: encoder(index(feature))

Fetch a batch and observe the features of a sample record. This will help in defining the keras preprocessing layers for training the tf.keras model.

ds_iter = iter(train_ds)
features, label = next(ds_iter)
{key: value.numpy()[0] for key,value in features.items()}
{'Age': 2,
 'Breed1': b'Tabby',
 'Color1': b'Black',
 'Color2': b'Cream',
 'Fee': 0,
 'FurLength': b'Short',
 'Gender': b'Male',
 'Health': b'Healthy',
 'MaturitySize': b'Small',
 'PhotoAmt': 4,
 'Sterilized': b'No',
 'Type': b'Cat',
 'Vaccinated': b'No'}

Choose a subset of features.

all_inputs = []
encoded_features = []

# Numeric features.
for header in ['PhotoAmt', 'Fee']:
  numeric_col = tf.keras.Input(shape=(1,), name=header)
  normalization_layer = get_normalization_layer(header, train_ds)
  encoded_numeric_col = normalization_layer(numeric_col)

# Categorical features encoded as string.
categorical_cols = ['Type', 'Color1', 'Color2', 'Gender', 'MaturitySize',
                    'FurLength', 'Vaccinated', 'Sterilized', 'Health', 'Breed1']
for header in categorical_cols:
  categorical_col = tf.keras.Input(shape=(1,), name=header, dtype='string')
  encoding_layer = get_category_encoding_layer(header, train_ds, dtype='string',
  encoded_categorical_col = encoding_layer(categorical_col)

Build, compile and train the model

# Set the parameters

# Convert the feature columns into a tf.keras layer
all_features = tf.keras.layers.concatenate(encoded_features)

# design/build the model
x = tf.keras.layers.Dense(32, activation="relu")(all_features)
x = tf.keras.layers.Dropout(0.5)(x)
x = tf.keras.layers.Dense(64, activation="relu")(x)
x = tf.keras.layers.Dropout(0.5)(x)
output = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(all_inputs, output)

tf.keras.utils.plot_model(model, rankdir='LR', show_shapes=True)


# compile the model
model.compile(optimizer=OPTIMIZER, loss=LOSS, metrics=METRICS)
# fit the model, epochs=EPOCHS)
Epoch 1/10

/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/ UserWarning: Input dict contained keys ['Age'] which did not match any model input. They will be ignored by the model.
  [n for n in tensors.keys() if n not in ref_input_names])

253/253 [==============================] - 4s 14ms/step - loss: 0.6169 - accuracy: 0.6042
Epoch 2/10
253/253 [==============================] - 4s 14ms/step - loss: 0.5634 - accuracy: 0.6937
Epoch 3/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5573 - accuracy: 0.6981
Epoch 4/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5528 - accuracy: 0.7087
Epoch 5/10
253/253 [==============================] - 4s 14ms/step - loss: 0.5512 - accuracy: 0.7173
Epoch 6/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5456 - accuracy: 0.7219
Epoch 7/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5397 - accuracy: 0.7283
Epoch 8/10
253/253 [==============================] - 4s 14ms/step - loss: 0.5385 - accuracy: 0.7331
Epoch 9/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5355 - accuracy: 0.7326
Epoch 10/10
253/253 [==============================] - 4s 15ms/step - loss: 0.5412 - accuracy: 0.7321

<tensorflow.python.keras.callbacks.History at 0x7f5c235112e8>

Infer on the test data

res = model.evaluate(test_ds)
print("test loss, test acc:", res)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/ UserWarning: Input dict contained keys ['Age'] which did not match any model input. They will be ignored by the model.
  [n for n in tensors.keys() if n not in ref_input_names])

109/109 [==============================] - 2s 15ms/step - loss: 0.5344 - accuracy: 0.7421
test loss, test acc: [0.534355640411377, 0.7420566082000732]