restgdf

lightweight async Esri REST client with optional GeoPandas extras

Python ≥ 3.11 | async-first | zero mandatory geo dependencies | pydantic-validated responses

import asyncio
from aiohttp import ClientSession
from restgdf import FeatureLayer

async def main():
    async with ClientSession() as session:
        layer = await FeatureLayer.from_url(
            "https://maps1.vcgov.org/arcgis/rest/services/Beaches/MapServer/6",
            session=session,
        )
        print(layer.name, layer.count)
        async for row in layer.stream_rows():
            print(row)
            break

asyncio.run(main())

pip install restgdf[resilience]

gpd.read_file(url, driver="ESRIJSON") does not account for max record count limitations, so large services get truncated at the server's maxRecordCount.

restgdf uses asyncio to read all features from a service, not just the first page, while letting you choose between a light-core install and an optional GeoPandas extra.

Installation

Requires Python ≥ 3.11 (3.9 reached EOL 2025-10-31; 3.10 reaches EOL 2026-10-31).

Install the lightweight core package when you want typed metadata, query helpers, crawl/auth utilities, or raw feature rows without pulling in pandas, geopandas, or pyogrio:

pip install restgdf

Base-install capabilities include:

• typed pydantic response models like LayerMetadata and CrawlReport
• FeatureLayer.from_url, .metadata, .count, and .get_oids()
• single-field get_unique_values() queries
• raw feature dictionaries via FeatureLayer.stream_features() / stream_rows() (deprecated row_dict_generator() still works)
• Directory crawling and ArcGISTokenSession authentication helpers

Install the geo extra for GeoDataFrame and pandas-backed workflows:

pip install "restgdf[geo]"

restgdf[geo] adds:

• FeatureLayer.get_gdf() / deprecated getgdf()
• FeatureLayer.sample_gdf() and head_gdf()
• FeatureLayer.fieldtypes
• pandas-backed helpers like get_value_counts() and get_nested_count()
• low-level restgdf.utils.getgdf helpers

Install the resilience extra for automatic retry with exponential back-off and configurable (per-service-root or per-host) rate limiting — useful for bulk/multi-server crawls:

pip install "restgdf[resilience]"

restgdf[resilience] adds stamina and aiolimiter. It changes no request behavior by itself — wrap your own aiohttp.ClientSession with restgdf.resilience.ResilientSession (configured via restgdf.ResilienceConfig) and pass the wrapped session as session= to FeatureLayer/Directory. See the polite bulk crawl recipe for a full worked example.

Treat the split above as the stable dependency boundary: geo-enabled environments should depend on restgdf[geo] explicitly. See MIGRATION.md for the 1.x → 2.0 and 2.x → 3.0 upgrade guides.

Light-core usage

import asyncio

from aiohttp import ClientSession

from restgdf import FeatureLayer


beaches_url = r"https://maps1.vcgov.org/arcgis/rest/services/Beaches/MapServer/6"


async def main():
    async with ClientSession() as session:
        beaches = await FeatureLayer.from_url(beaches_url, session=session)
        cities = await beaches.get_unique_values("CITY")

        first_rows = []
        async for row in beaches.stream_rows(data={"outFields": "CITY,STATE"}):
            first_rows.append(row)
            if len(first_rows) == 2:
                break

    return beaches.count, beaches.metadata.max_record_count, cities[:3], first_rows


count, max_record_count, cities, first_rows = asyncio.run(main())

print(count, max_record_count)
print(cities)
print(first_rows[0])

Streaming

FeatureLayer exposes ArcGIS pagination as three async generators so you can process millions of rows without buffering them in memory. The on_truncation knob controls what happens when the server caps a page at maxRecordCount: "raise" (default), "ignore" (log + continue), or "split" (bisect by object-id and retry, up to depth 32).

import asyncio

from aiohttp import ClientSession

from restgdf import FeatureLayer


zipcodes_url = "https://services.arcgis.com/P3ePLMYs2RVChkJx/ArcGIS/rest/services/USA_ZIP_Codes_2016/FeatureServer/0"


async def main():
    async with ClientSession() as session:
        oh = await FeatureLayer.from_url(
            zipcodes_url,
            where="STATE = 'OH'",
            session=session,
        )

        # 1. Feature dicts, one per row (base install)
        first_feature = None
        async for feature in oh.stream_features(on_truncation="split"):
            first_feature = feature
            break

        # 2. Per-page batches, preserving ArcGIS page boundaries
        page_sizes = []
        async for batch in oh.stream_feature_batches(order="request"):
            page_sizes.append(len(batch))
            if len(page_sizes) == 3:
                break

        # 3. GeoDataFrame chunks (requires `restgdf[geo]`; note that
        # stream_gdf_chunks does *not* accept on_truncation / order /
        # max_concurrent_pages — it yields in completion order).
        chunk_shapes = []
        async for chunk in oh.stream_gdf_chunks():
            chunk_shapes.append(chunk.shape)
            if len(chunk_shapes) == 2:
                break

    return first_feature, page_sizes, chunk_shapes


first_feature, page_sizes, chunk_shapes = asyncio.run(main())

See the streaming recipe for the full matrix of on_truncation, order, and max_concurrent_pages combinations on the iter_pages-based shapes (stream_features, stream_feature_batches, stream_rows).

GeoDataFrame workflows (restgdf[geo])

import asyncio

from aiohttp import ClientSession

from restgdf import FeatureLayer


beaches_url = r"https://maps1.vcgov.org/arcgis/rest/services/Beaches/MapServer/6"

zipcodes_url = "https://services.arcgis.com/P3ePLMYs2RVChkJx/ArcGIS/rest/services/USA_ZIP_Codes_2016/FeatureServer/0"


async def main():
    async with ClientSession() as session:
        beaches = await FeatureLayer.from_url(beaches_url, session=session)
        beaches_gdf = await beaches.get_gdf()

        daytona = await beaches.where("LOWER(City) LIKE 'daytona%'")
        daytona_gdf = await daytona.get_gdf()

        oh_zipcodes = await FeatureLayer.from_url(
            zipcodes_url,
            where="STATE = 'OH'",
            session=session,
        )
        oh_zipcodes_gdf = await oh_zipcodes.get_gdf()

    return beaches_gdf, daytona_gdf, oh_zipcodes_gdf


beaches_gdf, daytona_gdf, oh_zipcodes_gdf = asyncio.run(main())

print(beaches_gdf.shape)
# (243, 10)

print(daytona_gdf.shape)
# (83, 10)

print(oh_zipcodes_gdf.shape)
# (1026, 8)

Keyword arguments to FeatureLayer.get_gdf() are passed on to aiohttp.ClientSession.post; include query parameters like where and token in the data dict when needed.

Truncation now raises (PAGINATION-01). get_gdf() and stream_gdf_chunks() — the GeoDataFrame path — now inspect each page for exceededTransferLimit=true and raise restgdf.errors.PaginationError, mirroring the raw-feature streaming engine, instead of silently returning a GeoDataFrame that is missing rows. If a very large layer trips this, read it through the iter_pages-based stream_features / stream_rows shapes, which expose on_truncation="ignore" | "split" (the legacy geo chunk path does not take that knob).

Token authentication

Token helpers are available in the base install. The GeoDataFrame example below requires restgdf[geo] because it calls get_gdf().

import asyncio

from aiohttp import ClientSession

from restgdf import AGOLUserPass, ArcGISTokenSession, FeatureLayer


secured_url = "https://example.com/arcgis/rest/services/Secured/FeatureServer/0"


async def main():
    async with ClientSession() as base_session:
        token_session = ArcGISTokenSession(
            session=base_session,
            credentials=AGOLUserPass(
                username="my-username",
                password="my-password",
            ),
        )
        layer = await FeatureLayer.from_url(secured_url, session=token_session)
        return await layer.get_gdf()


secured_gdf = asyncio.run(main())

If you already have a token, you can pass it with token="..." or data={"token": "..."}. restgdf sends a token-bearing request as a POST with the token in the request body (never serialized into the URL query string), so the token does not leak into server or proxy access logs — this holds even for short requests that would otherwise use GET (AUTH-01). For fully header-based auth, use an ArcGISTokenSession, which sends the token via the X-Esri-Authorization header instead.

Typed responses

Typed responses are part of the base pip install restgdf surface.

Every response is a pydantic model. Attribute access replaces dict indexing, and model_dump(by_alias=True) round-trips back to ArcGIS camelCase:

import asyncio

from aiohttp import ClientSession

from restgdf import FeatureLayer


async def main():
    async with ClientSession() as session:
        fl = await FeatureLayer.from_url(beaches_url, session=session)
        md = fl.metadata                      # restgdf.LayerMetadata
        return md.name, md.max_record_count, md.model_dump(by_alias=True)


name, max_record_count, arcgis_dict = asyncio.run(main())

Need a plain dict during a transitional migration? Use restgdf.compat.as_dict(md). See MIGRATION.md for the 1.x → 2.0 and 2.x → 3.0 upgrade guides.

ArcGIS drift guarantees and limitations

• Strict envelopes (CountResponse, ObjectIdsResponse, TokenResponse) fail fast with RestgdfResponseError when required keys are missing or malformed.
• Permissive envelopes (LayerMetadata, FeaturesResponse, crawl/service payloads) still tolerate unknown extras and missing optional fields, but a top-level ArcGIS JSON error envelope ({"error": {...}}) now raises RestgdfResponseError instead of being mistaken for partial metadata or an empty feature page.
• Query helpers intentionally call aiohttp JSON decoding with content_type=None, so mislabeled JSON bodies such as text/plain still parse.
• Non-JSON/HTML bodies are not normalized: malformed query bodies still bubble the underlying JSON decoder error, and ArcGISTokenSession.update_token() still preserves aiohttp's native ContentTypeError behavior for HTML token pages.

Documentation

Full docs live at https://restgdf.readthedocs.io/ (hosted by Read the Docs).

Docs for humans and LLMs

Every page is published in three formats so you can feed it to a teammate or to a language model without any preprocessing:

Point your coding agent or RAG pipeline at llms-full.txt for the entire reference in a single file, or at llms.txt for a concise table of contents.

For contributors

• CONTRIBUTING.md — local setup, PR checklist, commit conventions, gate suite.
• ARCHITECTURE.md — module layout, exception taxonomy, logger hierarchy, config precedence, session ownership, streaming shapes, extras matrix.
• CHANGELOG.md — every user-visible change.
• MIGRATION.md — upgrading across 1.x → 2.0, 2.x → 3.0, and 3.0.x → 3.1.

Uses

• restgdf_api
• govgis_nov2023 — a bulk, multi-host crawl; see the polite bulk crawl recipe for the resilience + concurrency + User-Agent composition this kind of workload needs.