Telemetry

Starting with version 3.46.0.12, H2O-3 can send anonymous usage telemetry to help the team prioritize features and platforms. It is opt-in and off by default — nothing is sent unless you turn it on. It is also designed to be invisible when enabled: every send is fire-and-forget with a short timeout, so if the receiver is unreachable your code behaves exactly as if telemetry never ran — it never blocks, raises, or retries.

It is limited to operational and runtime characteristics, and is used only in aggregate to help H2O.ai operate, secure, support, maintain, and improve H2O-3 — understanding common deployment patterns, prioritizing compatibility and support efforts, and detecting abuse or malicious activity. It is not used for sales prospecting, lead identification, or user surveillance.

What is sent (when enabled)

  • One small ping when you start or connect to H2O (h2o.init() / h2o.connect() in Python or R, or a standalone java -jar h2o.jar / hadoop jar h2odriver.jar cluster). Depending on the client, major actions may also send one ping each — for example training, scoring, MOJO and model download, upload, import, parse, AutoML, and model save/load.

  • Each ping contains the H2O version, the client (python / r / jvm), the operating system, an ephemeral session ID regenerated on every start, and a timestamp. Depending on the action, a ping may also carry the algorithm name and coarse range buckets for counts such as rows, columns, durations, and sizes. These counts are reported as ranges rather than exact figures. One notable exception: a small cluster’s node count (1–16) is sent exactly, since 1-node vs 4-node is operationally meaningful; larger clusters are bucketed.

What is never sent

Code, dataset contents, prompts, model inputs or outputs, training data, prediction values, file paths or URLs, dataset or model names, column names, parameter values, hostnames, usernames, email addresses, precise location, or any other customer business data.

Source IP addresses are inherently visible to any HTTPS request; the receiver may use them transiently to derive a coarse geographic region and for network attribution, and does not persist or store them. The telemetry payload itself contains no location data.

Enabling telemetry

Telemetry is off by default. Turn it on with any of the following.

Python and R clients

  • Per session — pass telemetry=True (Python) or telemetry = TRUE (R) to h2o.init() or h2o.connect().

  • Persistent — use the setter h2o.set_telemetry() to change it and the getter h2o.telemetry_enabled() to read the current state. The setting applies immediately and is saved under ~/.h2oai/telemetry so later sessions honor it.

  • Config file — add a general.telemetry key to ~/.h2oconfig in your home directory:

    [general]
telemetry = true

Below is a simple example showing how to opt in persistently and read the current state.

R

Python

library(h2o)
h2o.init()

# Opt in (persisted across sessions)
h2o.set_telemetry(TRUE)

# Opt back out
h2o.set_telemetry(FALSE)

# Read the current state
h2o.telemetry_enabled()

import h2o
h2o.init()

# Opt in (persisted across sessions)
h2o.set_telemetry(True)

# Opt back out
h2o.set_telemetry(False)

# Read the current state
h2o.telemetry_enabled()

Standalone / Hadoop cluster (JVM)

A cluster started directly on the JVM (java -jar h2o.jar / hadoop jar h2odriver.jar) is also off by default. Enable it by explicitly setting the disable flag to false:

java -Dsys.ai.h2o.telemetry.disabled=false -jar h2o.jar

Turning it off / keeping it off

Since telemetry is off by default you normally don’t need to do anything. To turn it off after enabling it — or to force it off regardless of the settings above — use any of:

  • Clients — pass telemetry=False (Python) / telemetry = FALSE (R) for the current session, or h2o.set_telemetry(False) to persist it, or set general.telemetry = false in ~/.h2oconfig.

  • JVM cluster — -Dsys.ai.h2o.telemetry.disabled=true (the default, off, is also honored when the flag is unset).

  • Any environment — set DO_NOT_TRACK=1 (the cross-tool standard from consoledonottrack.com). This is a hard opt-out: it always wins over every enable setting above and is honored by the Python client, the R client, and the JVM server.

The receiver also honors the standard DNT: 1 (Do Not Track) and Sec-GPC: 1 (Global Privacy Control) request headers: any event arriving with either header set is dropped and never stored.