This document describes the detailed security threat model for Apache Iceberg Python. It is intended for maintainers and automated security triage.
Apache Iceberg Python is primarily a client library and implementation of the Iceberg table format and catalog interactions for Python applications and services. It is commonly embedded in larger systems that provide their own authentication, authorization, and credential management. Because of that deployment model, many bug classes that look security-relevant in the abstract are not actually security vulnerabilities in Iceberg Python itself.
This model is intended to answer:
This model is scoped to the Apache Iceberg Python repository itself:
It is not a general threat model for every Python service that embeds Iceberg Python.
In particular, it does not attempt to define the complete security model for:
Iceberg Python should:
Iceberg Python does not aim to be the primary enforcement point for:
That last point is worth stating plainly, because it is the one most often assumed. Iceberg Python does not guarantee that state is partitioned between catalog instances, sessions, or tenants within one interpreter. Caches, client objects, and auth state may be process-wide. A deployment that serves multiple principals from one process and needs them isolated must provide that isolation itself — by process, interpreter, or another boundary it controls. Reports that a process-wide structure is shared across catalogs or tenants are accepted as bugs worth fixing, not as vulnerabilities.
The operator configures the surrounding catalog, application, service, and storage integration around Iceberg Python. This role is trusted to choose endpoints, warehouses, storage integrations, and credentials.
The catalog control plane resolves tables and supplies metadata, locations, configuration, and delegated credentials to Iceberg Python. It may be implemented by a REST catalog server or another catalog implementation. Iceberg Python assumes this control plane is trusted and outside its primary security boundary.
The REST catalog client consumes catalog-provided metadata, configuration, and credentials. Client-side bugs in routing, caching, or reuse are real bugs and worth fixing, but they are security-relevant only where they expose a secret to an audience outside the process — not merely because state was shared between catalogs or sessions within it.
Applications and services embedding Iceberg Python are responsible for their own user-facing authorization boundaries unless Iceberg Python explicitly documents otherwise.
This role may already have legitimate power to write or replace table metadata, write or delete files, choose paths under an allowed warehouse or table location, and invoke destructive maintenance operations. If a report only shows a new way to achieve the same effect this role can already cause legitimately, it is usually not a security issue in Iceberg Python.
The following are generally treated as trusted operator or deployment inputs:
If a report depends on the attacker controlling those values directly, it is usually not a vulnerability in Iceberg Python itself.
Iceberg Python often accepts metadata locations, table properties, namespace properties, and related control-plane information from a catalog. By default, Iceberg Python treats those sources as trusted.
This means a malicious catalog supplying incorrect or malicious metadata is usually not an Iceberg Python vulnerability by itself.
In REST deployments, Iceberg Python may also accept service endpoints, configuration, and delegated storage access from the REST catalog server. By default, those are treated as trusted control-plane inputs unless Iceberg Python explicitly documents a stronger guarantee.
This means a malicious REST catalog server sending dangerous endpoints is usually not an Iceberg Python vulnerability by itself. It also means many credential-selection bugs are often correctness or specification issues rather than security boundary failures.
The major exception is secret exposure. If Iceberg Python surfaces credentials or secrets to a new audience that was not already trusted with them, that is security-relevant.
Object store permissions are enforced by the storage provider and the credentials the surrounding deployment chooses to hand to Iceberg Python. Iceberg Python is not the root authority for bucket- or object-level authorization.
The following categories are generally security-relevant in Iceberg Python when the report is credible and reproducible.
Examples include:
Sharing of credential-bearing state between catalogs or sessions inside a single process is not covered here; see Security Goals on process isolation.
Security issues exist when Iceberg Python is the documented enforcement point for a boundary and fails to hold it.
The qualifier matters. A boundary belongs to Iceberg Python only where this model or the project's documentation says it does. Where the surrounding catalog, service, or deployment is the enforcement point, a client-side gap is a bug rather than a vulnerability — even when the consequence looks alarming. In particular, isolation between principals sharing one process is not an Iceberg Python-owned boundary.
These categories may still be real bugs worth fixing, but they are not usually security vulnerabilities in Iceberg Python itself.
Examples include incorrect metadata handling, ambiguous matching semantics, and logic bugs that do not create a new trust-boundary violation.
Malformed-input crashes, runtime exceptions, and memory amplification are usually treated as robustness or hardening work rather than security issues in Iceberg Python itself.
Reports that require a malicious catalog or other external control-plane service are usually outside Iceberg Python's primary security boundary.
If the actor already has a legitimate capability that can cause the same harm, the new path is usually not a security issue.
A principal who can already write or replace table metadata, set table properties, or write and delete table data is operating inside the capability set described under Table writer or maintainer. A report that only shows a new route to an effect that principal can already produce legitimately is not a security vulnerability in Iceberg Python.
This covers, by default:
Containment of what a writer commits — that manifest entries and write paths stay inside the table's location — is enforced by the catalog, not by this client. A report showing that Iceberg Python follows a path the catalog accepted is a hardening opportunity here and a containment question there.
Availability-only findings are not treated as security vulnerabilities in Iceberg Python by default. This includes:
Iceberg Python is a client library; the surrounding service owns request admission, resource limits, and process isolation. These reports are accepted as robustness and hardening work, not as vulnerabilities, unless they also demonstrate a confidentiality or integrity consequence.
Bounds-checking gaps in native or compiled decode paths, including the Cython Avro decoders, are treated by default as parser hardening rather than as security vulnerabilities. Crashes, aborts, and read overruns that terminate the process fall under this rule.
This applies to the root cause, not to a theory about the consequence. A report that a bounds check is missing is hardening. A report that demonstrates specific memory contents reaching an attacker-observable channel is a different report, and is assessed on what it shows.
Iceberg Python loads configuration from documented default locations, including .pyiceberg.yaml resolved from the working directory, the user home directory, and PYICEBERG_HOME, together with PYICEBERG_-prefixed environment variables. This resolution order is public, documented behavior and follows the same pattern as configuration loading in widely used tooling across the ecosystem.
A report is not a vulnerability in Iceberg Python merely because it observes that configuration is discovered this way — for example, that a configuration file placed in a directory a user later works from will be read when Iceberg Python runs there. Choosing which directories to execute in, and which configuration files to trust on the local filesystem, belongs to the operator and to the surrounding environment, not to this library.
Establishing the provenance of a catalog or table is the reader's responsibility. Iceberg Python reads the metadata, manifests, and data files it is pointed at. It does not authenticate the authorship of that content, does not treat it as sandboxed input, and does not claim to.
A caller who does not know who wrote a table is in the same position as one calling an API endpoint whose operator they cannot identify: the decision to trust the source was made before the library was involved, and the library is not where it can be recovered.
Publicly published datasets are the common example, but they are only an example. The same reasoning applies to a metadata.json obtained out of band, a table registered by URL on a user's behalf, a warehouse path an operator cannot vouch for, or any other route by which content of unestablished authorship reaches a read. What matters is whether provenance was established, not how the content arrived.
Reports whose precondition is that provenance was never established are therefore outside Iceberg Python's primary security boundary by default. Where a deployment needs that guarantee, it belongs upstream of this library — in the catalog that vends the table, or in whatever admits a dataset into the deployment.
A scanner targeting Iceberg Python should treat a finding as higher-confidence only if it plausibly shows one of the following:
A finding should be downgraded or rejected by default if it instead depends primarily on: