Skip to main content

Getting the Data

How the data set is delivered, from subscribing to reading the files. The Quickstart walks through the same steps for one day.

Access​

The data set is a free product on AWS Data Exchange. Subscribe on the product page; we review each request, which takes a few days. Subscribing means agreeing to the Data Subscriber Agreement (DSA). The data itself is licensed under CC BY-NC-SA 4.0.

Once approved, the data set appears in the AWS Data Exchange console under Entitled data, in the us-east-1 region, as pureskillgg-csgo-production-dataexchange-csds-0. Its Data set ID is what the export functions below take.

Revisions​

  • One revision a day. A revision opens at 00:00 UTC and collects every match processed that UTC day. It is published when the next one opens. Its comment is the moment it was created, which is how the export functions pick revisions by date.
  • Processed, not played. A match sits in the revision for the day it was processed, which can differ from the day it was played. Before 2026-09-07, a revision that failed to close carried its matches into the next one.
  • Pruned once a year. Revisions more than about a year old are revoked and emptied in a batch about once a year, so the window runs from one to two years. Revoked revisions still appear in the revision list but can't be exported. Revision Stats lists what each revision holds.

Exporting​

Data Exchange delivers a revision by writing its objects into an S3 bucket you own. The dsdk library wraps the calls:

FunctionWhat it does
export_multiple_adx_dataset_revisions_to_s3(bucket, data_set_id, start_date=…, end_date=…)Exports every revision created from start_date up to, not including, end_date.
export_single_adx_dataset_revision_to_s3(bucket, data_set_id, revision_id)Exports one revision; the latest one if revision_id is left out.
enable_auto_exporting_adx_dataset_revisions_to_s3(bucket, data_set_id)Exports each new revision as it is published, until disable_auto_exporting_adx_dataset_revisions_to_s3(data_set_id).
download_adx_dataset_revision(root_path, data_set_id, revision_id)Downloads a revision straight to disk, one object at a time, and unzips zipped matches. Slow for a whole day.
extract_adx_zips(root_path, delete_zips=False)Unzips every match zip under a folder into the loose layout, for revisions you exported to S3 and synced down.
get_adx_dataset_revisions(data_set_id, start_date=…, end_date=…)Lists the revisions created in a date range.

They use boto3's credentials and default region, so set the region to us-east-1. Each export keeps every object's own key, as the File layout shows; pass prefix= to put the keys under a folder. In the console, a revision's export to Amazon S3 does the same when its key pattern is ${Asset.Name}.

Permissions. The AWS identity that exports needs dataexchange:GetDataSet, ListDataSetRevisions, GetRevision, CreateJob, StartJob and GetJob, plus ListRevisionAssets to download to disk, and CreateEventAction, ListEventActions and DeleteEventAction to turn auto-export on and off. It also needs s3:PutObject on the bucket. Auto-export writes as Data Exchange itself, so the bucket's policy must allow that:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": { "Service": "dataexchange.amazonaws.com" },
      "Action": ["s3:PutObject", "s3:PutObjectAcl"],
      "Resource": "arn:aws:s3:::your-bucket/*",
      "Condition": {
        "StringEquals": { "aws:SourceAccount": "your-account-id" }
      }
    }
  ]
}

Cost. The Cost FAQ has the sizes and how to keep the bill down. A bucket in us-east-1 avoids transfer charges between regions.

File layout​

Newer revisions deliver each match as one zip; older ones deliver its files loose. The Changelog records the first revision with zips once they start. Both unpack to the same layout:

csds/
  2026/10/07/            the UTC day the match was processed
    <id>.zip             newer revisions: the whole match, one object
    <id>/                one folder per match, named by its public id
      csds               the index object: gzipped JSON
      bomb_action        one Parquet file per channel
      bomb_defuse
      …                  55 channels in all
  • A match zip is csds/YYYY/MM/DD/<id>.zip. Its entries are the match's files under their full keys (csds/YYYY/MM/DD/<id>/csds, csds/YYYY/MM/DD/<id>/bomb_action, …), stored without recompression, so each is byte for byte the loose file. Unzipping it in the folder that holds csds/ rebuilds the layout above. dsdk's extract_adx_zips(root_path) does that for every zip under a folder, and download_adx_dataset_revision does it for you. Any zip tool or Python's zipfile opens one, and a single channel can be read without unpacking the rest.
  • Keys of the loose files are csds/YYYY/MM/DD/<id>/<channel>, with no file extensions. Keep the whole key when you download or unzip: the libraries find a match's files by it.
  • Channel files are Apache Parquet, compressed with ZSTD since 2026-07-08 (GZIP before). Any Parquet reader opens them, for example pandas.read_parquet(path).
  • The index object is JSON compressed with gzip: json.loads(gzip.decompress(open(path, "rb").read())).
  • Object counts. A zipped match is one object holding 56 files, or fewer in matches from before the newer channels (convar, team_info, pause_state, smoke_voxel, world_object, hostage_vector, hostage_state, player_command, player_subtick_input, map_ping, player_round_stats, round_facts and vote) were added. A loose match is 43 objects since 2026-08-04 and 31 before; the 2026-08-03 revision holds both, and so does the revision in which zips start. Read the index rather than assuming a channel list.

The index object​

FieldWhat it holds
idThe match's public id, also its folder name.
keyThe index object's own key.
createdAtWhen the match was processed (UTC); its date is the one in the key.
matchDateWhen the match was played, cut to the minute. For a match uploaded by hand, when it was processed.
platformsteam, faceit or unknown. Matches uploaded by hand are unknown.
matchTypeThe match's type, such as competitive.
gamecsgo, for every match, CS2 included.
channelsOne entry per channel, below.
sharecode, demoIdredacted; see PII Removal.
metadata, context, timings, jobId, type, redactedPipeline bookkeeping.

Each entry in channels has:

  • channel, the channel's name, and key, its file's key;
  • category: header, single_event, multi_event, telemetry or player_info, as in the CSDS Spec;
  • events, the game events that write rows into it;
  • columns, one entry per column with name, type, nullable, origin, dependents, mergeKeys and comment, as the spec's How to read this describes. Skip a column whose origin ends in -deleted: the file no longer stores it.
  • available, on eight channels whose events a demo may not carry (bot_takeover, bullet_damage, player_command, player_inputs, player_sound, player_subtick_input, rank_update and round_mvp). false means the demo didn't carry the event, so the channel is empty for that reason; true means it did, though the channel can still be empty when nothing happened. In the revision of 2026-10-07, player_inputs and player_sound were false in 88 of 89 matches. From parser 5.7.0 (2026-10-12), player_inputs is true on every kind of demo, and player_sound on every demo that carries the game's sound messages (see Known Flaws).

Match identity and dates​

  • A match's id is new each time a demo is processed. A demo processed twice, for example one uploaded by two users, appears twice with different ids. Find duplicates by comparing header columns such as map_name, server_name, build_num and the final scores.
  • Which date to use. header.match_date is when the match was played, except for matches uploaded by hand, where it is when they were processed. The date in the key is always when the match was processed.