Cloud Delivery is in limited availability. To request access, reach out to your sales or account manager.
Overview
Coin Metrics publishes Cloud Delivery datasets to an S3 bucket that Coin Metrics owns and operates. You read objects from that bucket directly using your own AWS account, with standard tooling such as the AWS CLI, the AWS SDKs, DuckDB, Spark, Athena, or Glue. Nothing is copied into your account unless you copy it yourself. This page covers granting your account access, the layout of the bucket, and reading the files. For which datasets exist and how far back they go, see Cloud Delivery.Get access
Access is granted per customer through cross-account IAM. Anonymous and public access are not supported.What to provide
Send your account manager the following.
Coin Metrics grants read-only access to the prefixes covering your entitled datasets. You receive
s3:GetObject on those prefixes and s3:ListBucket scoped to them. Write, delete, and ACL operations are never granted.
Requester Pays
The bucket runs in Requester Pays mode. Coin Metrics pays to store the data. You pay the S3 request and data-transfer charges for what you read, billed to your own AWS account. This has a practical consequence: every request must declare that you accept the charges, or S3 rejects it with403 Access Denied. How you declare it depends on the tool.
- AWS CLI
- boto3
- s3fs / PyArrow
Understand the layout
Objects use Hive-style partitioning, where each directory level is akey=value pair. Query engines including Athena, Glue, Spark, DuckDB, and PyArrow discover these partitions automatically and skip prefixes that cannot match your filters.
A trades object looks like this.
a-z0-9 collapsed to an underscore. An exchange such as binance.us therefore appears as exchange=binance_us.
Path templates
Levels differ by dataset. The order is always coarsest first, so filtering on an early level eliminates the most work.
Two shapes are worth noting. Reference rates carry no
market-type or exchange level, because the dataset is keyed by asset rather than by market. Reference data carries no dt level at all, so each of its four objects is overwritten in place every day.
Partition values
Files
Each partition holds Parquet files covering that day for that slice of the market universe. Filenames encode the markets requested and the date range covered, for examplecoinbase-all-spot-2026-07-28-2026-07-29.parquet. Treat filenames as opaque. Read every .parquet object under a partition prefix rather than reconstructing a name, since names can change as coverage expands.
A day on which a market produced nothing is skipped rather than written as an empty file, so a missing partition means no activity rather than missing data.
File format
Every object is an Apache Parquet file with Snappy compression applied inside the file, which is the Parquet default and is handled transparently by every Parquet reader. There is no outer.gz wrapper, so files end in .parquet and can be opened directly.
Numeric fields are stored as native Parquet types rather than the JSON strings the API returns. Timestamps are stored as nanosecond epoch values in UTC.
Read the data
Point any Parquet reader at a prefix. The examples below read one day of Coinbase spot trades.- AWS CLI
- pandas / PyArrow
- DuckDB
- Spark
Reading efficiently
Because you pay per request and per byte transferred, how you scan matters.- Filter on partition keys first. A query filtered to one exchange and one date reads a handful of objects. The same query without those filters lists and reads the whole dataset.
- Point readers at the deepest prefix you can. Passing the full
dt=prefix avoids listing every exchange. - Select only the columns you need. Parquet is columnar, so projecting a few columns transfers far less than reading whole rows.
- Avoid repeated full-bucket listings. Listing calls are billed like any other request.
Limits and notes
- You pay for reads. Requester Pays means request and transfer charges land on your AWS account. Broad scans across many exchanges and dates can be expensive.
- The data is read-only. Write, delete, and ACL permissions are never granted.
- Reference data has no history. Its four objects are overwritten daily and carry no date partition. Keep your own copies if you need point-in-time catalog history.
- Order book files are large. Full-book snapshots at raw granularity produce the biggest objects in the bucket. Filter tightly and expect long transfers.
- File names are not a stable interface. Read every
.parquetobject under a partition rather than depending on a naming pattern. - Encryption and transport. Objects are encrypted at rest with server-side encryption, and the bucket policy requires TLS.
Related
- Cloud Delivery: datasets, coverage, and the update schedule.
- Snowflake: the same data as tables in your warehouse.
- AWS documentation on Requester Pays buckets

