How the Data Fits Together

How a match's channels relate: what a row is, how time and rounds are counted, how to follow a player, and how to join an event to where the players were. The numbers on this page were measured on the 89 matches published on 2026-10-07. Enums says what the codes mean, and the CSDS Spec lists every column.

What a row is

Each channel has a category, shown in its heading in the spec:

Category	One row is	Channels
header	The match.	header
player_info	One player in one round, taken when the round's freeze time ends.	player_info, player_personal
telemetry	Something on one tick.	player_vector and player_status (one living player), tick (the tick), grenade_vector and world_item_vector (one object that moved), molotov_fire, player_inputs
single_event	One game event.	player_death, player_hurt, weapon_fire, round_end and the rest
multi_event	One of several related events, named in event_type.	bomb_action, bomb_state, grenade_state, molotov_state, round_state and the rest

Every row except the header's carries round, and every event and telemetry row carries tick.

Time

• tick counts the game's ticks: 64 a second (header.tick_rate).
• second is tick / 64.
• The tick channel has one row for every tick, with no gaps.

Rounds

• Rounds count from 1. Warmup, knife rounds and drawn rounds are removed before publication, so round 1 is the first live round and the numbers run on without gaps.
• round_end has one row per round: who won in winner_team_code (2 Terrorists, 3 Counter-Terrorists) and why in win_reason_code.
• header.max_rounds is the regulation length: 24, or 16 in Wingman. Rounds above it are overtime.
• round_state marks the moments within each round, such as round_freeze_end, buytime_ended, bomb_planted and round_end, in event_type.

Players

• Follow a player with player_id_fixed. It numbers the players of a match from 1 and stays the same for each player all match. Bots get their own number, and player_personal.is_bot says which they are.
• player_personal.steam_id is a letter per player (A, B, C …) that also stays the same all match. Neither it nor player_id_fixed means anything in another match; see Known Flaws on linking players across matches.
• player_id is the game's own id. Joins between channels use it with tick, as below.

Sides

• player_info.team_code gives each player's side in each round: 2 Terrorists, 3 Counter-Terrorists.
• Teams switch sides at the half, after round max_rounds / 2: every player in every regulation match we checked was on the other side in the next round.
• The header columns starting t_starters_ and ct_starters_ describe the team that started as Terrorists and the team that started as Counter-Terrorists, through the side swaps.

Where the state lives

player_vector and player_status hold one row per living player per tick, keyed on (tick, player_id), and the two share those keys exactly. A dead player has no rows for the rest of the round.

Channel	Holds
player_vector	Position (x_pos, y_pos, z_pos), view angles (theta_ang, phi_ang), the weapon in hand (weapon_code), ammo, recoil and crouching. Velocity and movement angle are computed on load, as the Quickstart shows.
player_status	Health, armor, money, the inventory (inv_*), equipment values, the callout the player is in (place_name), and flags such as is_scoped and has_defuser.

Positions and View Angles explains the coordinates. player_status can be empty in a match; one of the 89 matches of 2026-10-07 had no rows in it.

Joining events to positions

Most event channels already carry where their players were. A channel with a player_id usually has player_x_pos, player_theta_ang, player_weapon_code, player_team_code and the like, copied from player_vector; an attacker_id or assister_id brings the same columns for that player. grenade_bounce, player_connect, player_disconnect, player_footstep and round_mvp carry none, so check a channel's columns in the CSDS Spec.

• Most channels copy from the row on the event's own tick.
• player_death, player_hurt, bullet_damage, bomb_action, item_dropped, player_chat and rank_update copy from the latest row up to 3 ticks before the event, because the player may already be dead on the event's tick.
• Where no row is close enough, the columns are null. On 2026-10-07 that was 0.2% of victims and 0.3% of attackers in player_death, 0.3% of weapon_fire rows, and 63% of assisters, who are often dead by the time of the kill.

To bring in a column the event doesn't carry, join the same way. This adds each killer's money at the moment of the kill:

import pandas as pd

deaths = loader.get_channel({"channel": "player_death"})
status = loader.get_channel(
    {"channel": "player_status", "columns": ["tick", "player_id", "money"]}
)

kills = deaths.dropna(subset=["attacker_id"]).astype({"attacker_id": "int64"})
money = status.rename(columns={"player_id": "attacker_id", "money": "attacker_money"})
money = money.astype({"attacker_id": "int64", "tick": "int64"})

kills = pd.merge_asof(
    kills.sort_values("tick"),
    money.sort_values("tick"),
    on="tick",
    by="attacker_id",
    direction="backward",
    tolerance=3,
)
print(kills[["round", "tick", "weapon_name", "attacker_money"]].head())

loader is a match loaded as in the Quickstart.