Tensorflow 2.0 - tf.estimator.DNNClassifier training for big datasets

I'm trying to train a DNNClassifier

    labels = ['BENIGN', 'Syn', 'UDPLag', 'UDP', 'LDAP', 'MSSQL', 'NetBIOS', 'WebDDoS']

    # Build a DNN
    classifier = tf.estimator.DNNClassifier(
    feature_columns=feature_columns,
    hidden_units=[30, 10],
    n_classes=len(labels),
    label_vocabulary=labels)

    def input_fn(features, labels, training=True, batch_size=32):
       '''
       An input function for training or evaluating
       '''
       # Convert the inputs to a Dataset.
       dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
       # Shuffle and repeat if you are in training mode.
       if training:
          dataset = dataset.shuffle(1000).repeat()
       return dataset.batch(batch_size)

    # Train the model
    classifier.train(
    input_fn=lambda: input_fn(train_features, train_label, training=True),
    steps=5000)

The training works fine until a bigger dataset is used

train_features.shape
>>> (15891114, 20)
train_label.shape
>>> (15891114,)

I'm using Google Colaboratory and my session crashes for exceeding RAM use (12GB of RAM) as soon as the training starts

WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python

/ops/resource_variable_ops.py:1666: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/training/training_util.py:236: Variable.initialized_value (from tensorflow.python.ops.variables) is deprecated and will be removed in a future version.
Instructions for updating:
Use Variable.read_value. Variables in 2.X are initialized automatically both in eager and graph (inside tf.defun) contexts.
INFO:tensorflow:Calling model_fn.
WARNING:tensorflow:Layer dnn is casting an input tensor from dtype float64 to the layer's dtype of float32, which is new behavior in TensorFlow 2.  The layer has dtype float32 because it's dtype defaults to floatx.

If you intended to run this layer in float32, you can safely ignore this warning. If in doubt, this warning is likely only an issue if you are porting a TensorFlow 1.X model to TensorFlow 2.

To change all layers to have dtype float64 by default, call `tf.keras.backend.set_floatx('float64')`. To change just this layer, pass dtype='float64' to the layer constructor. If you are the author of this layer, you can disable autocasting by passing autocast=False to the base Layer constructor.

WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/adagrad.py:106: calling Constant.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.

Before the training starts only 1GB of RAM is used, but as soon as the training starts RAM is rapidly saturated.

I'm making it work by providing chunks of the dataframe to train/evaluate the model.

Still, it is not clear to me why the RAM is saturated when I provide the whole dataframe for training or evaluation of the Estimator.