Apache PredictionIO Is Retired: Status and Alternatives on AlmaLinux 9
Apache PredictionIO was retired in 2020 and CentOS 7 is end-of-life. Here is what that means and how to replace it with Spark MLlib and MLflow on AlmaLinux 9.
If you came here to install Apache PredictionIO on CentOS 7, the short answer is: do not start a new deployment with it. Both the software and the operating system have been retired. This article explains the current status of PredictionIO, what your options are if you already run it, and how to build the same kind of service — collect events, train a model, answer prediction queries over HTTP — on AlmaLinux 9 or Rocky Linux 9 with projects that are still maintained.
Status of Apache PredictionIO
Apache PredictionIO was a machine learning server for building and deploying predictive services. It became an Apache top-level project in October 2017, was retired in September 2020, and its move to the Apache Attic was completed in April 2021. The project website now opens with a retirement notice.
The Apache Attic page for PredictionIO lists what remains: the website, mailing-list archives, the source repository and the release downloads on the Apache archive server, all read-only. In practice this means:
- No new releases, bug fixes or security patches.
- No support for newer versions of the components it depends on. A full installation bundled Apache Spark and MLlib, HBase, Akka HTTP and Elasticsearch, and the last release was built against versions of those that are themselves old now.
- No community to ask. The mailing lists are closed.
CentOS 7, the platform older tutorials used, reached end of life in June 2024. Its successors for new servers are AlmaLinux 9, Rocky Linux 9 and CentOS Stream 9.
If you still run PredictionIO
An existing installation keeps working, but it is an unpatched Java stack that accepts data over HTTP. Until you migrate:
- Do not expose the event server or the engine's query port to the internet. Bind them to a private address or restrict them with
firewalldto the application servers that need them. - Export your event data. The events (user actions, item properties) are the valuable part and are independent of PredictionIO; the trained models can be rebuilt from them.
- Keep the engine template's source code. It documents which algorithm and parameters you used, which is what you will re-implement.
What replaces it
There is no single drop-in successor. PredictionIO combined three jobs, and each maps to a maintained tool:
| PredictionIO part | What it did | Maintained replacement |
|---|---|---|
| Event Server | Collected events over a REST API and stored them | Your application's own database, or an event log you already operate |
| Engine and templates | Trained models with Spark MLlib | Apache Spark MLlib used directly, or any Python ML library |
| Engine server | Answered prediction queries over HTTP | A model-serving layer such as MLflow's model server |
Apache Spark MLlib
PredictionIO's engines were wrappers around Spark's machine learning library, and that library is alive and actively released. MLlib provides classification, regression, clustering and collaborative filtering (ALS, the algorithm behind the popular recommendation template), plus pipelines and model persistence. The DataFrame-based spark.ml API is the primary one; the older RDD-based spark.mllib API that PredictionIO templates used is in maintenance mode. See the MLlib guide for the current API.
MLflow for packaging and serving
MLflow is an open-source platform that records training runs, stores the resulting models and can serve a stored model as an HTTP endpoint, which covers the "deploy the engine as a web service" step. Its model server accepts POST requests at /invocations and returns predictions.
A minimal replacement on AlmaLinux 9 or Rocky Linux 9
The following sets up MLflow in a Python virtual environment under the deploy user. It is a starting point for experiments, not a hardened production layout.
sudo dnf install -y python3 python3-pip
python3 -m venv ~/mlenv
source ~/mlenv/bin/activate
pip install --upgrade pip
pip install mlflow
Train a model with the library of your choice and log it with MLflow from your training script; each run gets a run ID. To serve a logged model on port 5000, replace <run_id> with that ID:
mlflow models serve -m runs:/<run_id>/model -p 5000
Query it the way you used to query a PredictionIO engine, with a JSON POST request. The shape of inputs depends on your model:
curl http://127.0.0.1:5000/invocations -H "Content-Type:application/json" --data '{"inputs": [[1, 2], [3, 4], [5, 6]]}'
Keep the port bound to localhost or a private network and put your application, or an authenticated reverse proxy, in front of it. For Spark-based training, install Spark from the Apache Spark downloads page and follow its documentation for the Java version it requires, rather than reusing the versions from an old PredictionIO guide.
Migration checklist
- Export the events from the old installation and load them into a store you control.
- Re-implement the training step with
spark.mlor a Python library, using the old engine template as the specification. - Compare predictions from the new model with the old engine on the same inputs.
- Serve the new model, switch your application's queries to the new endpoint, then shut the old stack down.
Frequently asked questions
Can I still download and install PredictionIO?
The old releases remain on the Apache archive server, so technically yes. You would be installing unmaintained software on top of outdated versions of Spark, HBase and Elasticsearch, with no security fixes. That is acceptable for studying the code, not for a service that handles real data.
Is there a maintained fork?
The Apache project itself points to none; the Attic page only lists read-only archives. Evaluate any third-party fork you find on its own release history and activity before trusting it.
Do I need Spark at all for a small data set?
Often not. Spark pays off when the data does not fit comfortably on one machine. For a few million events, a single-node Python library on a VPS is simpler to operate, and you can still serve the result through the same kind of HTTP endpoint.
What happened to the engine templates?
They were tied to PredictionIO's own framework and stopped being developed with it. The algorithms inside them, such as ALS for recommendations, are standard and available directly in Spark MLlib.