This is a bare-bones, minimal example for building an application on the
htmengine framework. There are two important steps that must be done
in order to use htmengine: Configure the application, and setup the
database. Once you have completed the following instructions, you will have
a running htmengine application. You may send data to the
graphite-compatible custom metrics interface, or send data directly into the
AMQP exchange, and receive real-time results from the model results exchange,
or by polling the database.
You will see in several instances the use of the word "skeleton". That is the name of the app in this demonstration. In all cases, you should replace "skeleton" with the name that you've chosen for your application. This value is used in several places to segment resources by application, so that multiple applications may use the same mysql and rabbitmq instance.
First, this assumes you have a fresh checkout of the numenta-apps repository
at https://github.com/numenta/numenta-apps. Clone that repository, and install
htmengine in development mode:
cd numenta-apps/htmengine
python setup.py develop --user
You'll also need to install nta.utils, which is a dependency of htmengine:
cd numenta-apps/nta.utils
python setup.py develop --user
For example:
mysql -u root --execute="CREATE DATABASE skeleton"
Inside conf/application.conf, you will find a section called [repository],
edit the db, host, user, passwd, and port values to match the
database created in Step 1.
Elsewhere in conf/application.conf, replace "skeleton" with the name of your
application. That includes results_exchange_name under [metric_streamer],
and queue_name in [metric_listener].
Inside conf/model-checkpoint.conf, specify an absolute path on your
filesystem for permanent storage of model checkpoints in root of [storage].
Inside conf/model-swapper.conf, replace "skeleton" with the name of your
application in results_queue, model_input_queue_prefix, and
scheduler_notification_queue of the [interface_bus] section.
Inside conf/supervisord.conf, specify the path to the configuration directory
in the environment value of the [supervisord] section. Specify the path to
the supervisord-base.conf bundled with the original htmengine installation.
For example, if you cloned the numenta-apps repository at
/Users/user/numenta-apps, the value should be
/Users/user/numenta-apps/htmengine/conf/supervisord-base.conf.
APPLICATION_CONFIG_PATH must be set, and include the path to the config files
mentioned in #2. For example, if you're in the root of this project:
export APPLICATION_CONFIG_PATH=`pwd`/conf
Again, from the root of this project:
python repository/migrate.py
Again, from the root of this project:
mkdir -p logs
supervisord -c conf/supervisord.conf
At this point, the core htmengine services are running. You can see the
supervisor status at http://localhost:9001/ or by running
supervisorctl status.
A sample send_cpu.py script is included demonstrating how to send data into
htmengine for processing using the message_bus_connector API.
python send_cpu.py
send_cpu.py will run indefinitely, sending cpu percent samples every 5
seconds into the custom metrics queue for processing by htmengine. In practice
you should use something like nohup to keep this process running in the
background:
nohup python send_cpu.py > send_cpu.stdout 2> send_cpu.stderr < /dev/null &
If you look in the metric_data MySQL table, you should start to see new
records come in.
For example:
$ mysql -u root skeleton --execute="select * from metric_data order by rowid desc limit 5"
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
| uid | rowid | timestamp | metric_value | raw_anomaly_score | anomaly_score | display_value |
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
| 4258abfc6de947609f821095471dd0a2 | 94 | 2015-06-13 04:01:55 | 5.6 | NULL | NULL | NULL |
| 4258abfc6de947609f821095471dd0a2 | 93 | 2015-06-13 04:01:50 | 7.4 | NULL | NULL | NULL |
| 4258abfc6de947609f821095471dd0a2 | 92 | 2015-06-13 04:01:45 | 4.1 | NULL | NULL | NULL |
| 4258abfc6de947609f821095471dd0a2 | 91 | 2015-06-13 04:01:40 | 3.7 | NULL | NULL | NULL |
| 4258abfc6de947609f821095471dd0a2 | 90 | 2015-06-13 04:01:35 | 4.5 | NULL | NULL | NULL |
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
Notice that only uid, rowid, timestamp, and metric_value are populated.
The next step will be to explicitly create a model for the cpu_percent metric
so that raw_anomaly_score and anomaly_score may be populated with anomaly
scores.
python create_cpu_percent_model.py
Then, check the database again, and you should see anomaly scores:
$ mysql -u root skeleton --execute="select * from metric_data order by rowid desc limit 5"
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
| uid | rowid | timestamp | metric_value | raw_anomaly_score | anomaly_score | display_value |
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
| 4258abfc6de947609f821095471dd0a2 | 344 | 2015-06-13 04:23:11 | 10.5 | 0.075 | 0.344578258 | 1000 |
| 4258abfc6de947609f821095471dd0a2 | 343 | 2015-06-13 04:23:06 | 9.6 | 0.05 | 0.184060125 | 1000 |
| 4258abfc6de947609f821095471dd0a2 | 342 | 2015-06-13 04:23:01 | 8.4 | 0.05 | 0.11506967 | 1000 |
| 4258abfc6de947609f821095471dd0a2 | 341 | 2015-06-13 04:22:56 | 9.4 | 0.05 | 0.080756659 | 1000 |
| 4258abfc6de947609f821095471dd0a2 | 340 | 2015-06-13 04:22:51 | 7.6 | 0 | 0.044565463 | 1000 |
+----------------------------------+-------+---------------------+--------------+-------------------+---------------+---------------+
You can query the metric table to get status on the model. You'll know
everything is working when status is 1:
$ mysql -u root skeleton --execute="select uid, name, description, status from metric where name = 'cpu_percent'"
+----------------------------------+-------------+---------------------------+--------+
| uid | name | description | status |
+----------------------------------+-------------+---------------------------+--------+
| 4258abfc6de947609f821095471dd0a2 | cpu_percent | Custom metric cpu_percent | 1 |
+----------------------------------+-------------+---------------------------+--------+
Additionally, you can consume the anomaly results in near realtime with the
included consume_realtime_results.py script.
$ python consume_realtime_results.py
Handling 1 model result(s) for 4258abfc6de947609f821095471dd0a2 - cpu_percent
4258abfc6de947609f821095471dd0a2 [{u'rowid': 749, u'rawAnomaly': 0.0, u'anomaly': 0.04456546299999997, u'ts': 1434174392.0, u'value': 8.0}]
Handling 1 model result(s) for 4258abfc6de947609f821095471dd0a2 - cpu_percent
4258abfc6de947609f821095471dd0a2 [{u'rowid': 750, u'rawAnomaly': 0.0, u'anomaly': 0.04456546299999997, u'ts': 1434174397.0, u'value': 9.6}]
Handling 1 model result(s) for 4258abfc6de947609f821095471dd0a2 - cpu_percent
4258abfc6de947609f821095471dd0a2 [{u'rowid': 751, u'rawAnomaly': 0.0, u'anomaly': 0.04456546299999997, u'ts': 1434174402.0, u'value': 7.3}]
Handling 1 model result(s) for 4258abfc6de947609f821095471dd0a2 - cpu_percent
4258abfc6de947609f821095471dd0a2 [{u'rowid': 752, u'rawAnomaly': 0.0, u'anomaly': 0.04456546299999997, u'ts': 1434174407.0, u'value': 8.1}]
Handling 1 model result(s) for 4258abfc6de947609f821095471dd0a2 - cpu_percent
4258abfc6de947609f821095471dd0a2 [{u'rowid': 753, u'rawAnomaly': 0.0, u'anomaly': 0.04456546299999997, u'ts': 1434174412.0, u'value': 7.0}]
...
See webapp.py for a minimal web service implementation that implements the
above described steps in the form of HTTP calls.
First, run the web service:
python webapp.py
Then, to create a model, send a PUT request to the web service on port 8080.
The URI represents the metric name, and you may optionally specify model params
in the request body. Using curl on the command line:
$ curl http://localhost:8080/load_average -X PUT -d '{"min":0, "max":12}' -i
HTTP/1.1 201 Created
Content-Type: text/html
Date: Sat, 13 Jun 2015 21:01:40 GMT
Transfer-Encoding: chunked
Created 6e1f199a74274c5cbf9443f4ab4ad94e
Note you may create the model at any time during the process. For example,
you may send data (described below) over an extended period of time, and then
trigger the creation of the model at a later time. In the example above, the
min and max is included, but that value may not be known. Sometimes it is
better to omit the min and max and let htmengine choose a value based on the
history of a metric.
To send data, send a POST request to the same URL you used to create the
model. The body of the request must consist of two values: metric value and
timestamp separated by a single space. The example below sends load average
data using curl in an infinite loop at 5 second intervals:
$ while true; do while true; do echo "`(uptime | awk -F'[a-z]:' '{ print $2}' | awk '{print $1}' | tr -d ',')` `date +%s`" | curl http://localhost:8080/load_average -X POST -d @-; sleep 5; done; done
Saved /load_average 1.620000 @ 1434229721
Saved /load_average 1.730000 @ 1434229726
Saved /load_average 1.590000 @ 1434229731
Saved /load_average 1.540000 @ 1434229736
Saved /load_average 1.420000 @ 1434229741
To retrieve data, GET data from the same original URL:
$ curl http://localhost:8080/load_average
/load_average 1.7 1434229625 0.0
/load_average 1.72 1434229628 0.0
/load_average 1.72 1434229629 0.0
/load_average 1.72 1434229630 0.0
/load_average 1.71 1434229704 0.0
/load_average 1.73 1434229709 0.0
/load_average 1.62 1434229721 0.0
/load_average 1.73 1434229726 0.0
/load_average 1.59 1434229731 0.0
/load_average 1.54 1434229736 0.0
/load_average 1.42 1434229741 0.0
The right-most column in the response is the anomaly score. The first three columns are the original data.
Disclaimer this web app is only a simple demonstration wrapping htmengine in a minimal web service and is not production-ready.
This demonstration represents only the minimal amount of work to bring an
application online using htmengine. Numenta have done a substantial amount
of work in making much of the concepts demonstrated here robust and suitable
for a production workload in our showcase applications Grok, and Taurus. If
you're interested in learning more, please see
Grok,
Taurus Engine,
and Taurus Metric Collectors
Thanks!