Skip to main content

CSDS Changelog

Every change to the published CS2 CSDS data since it began on 2023-11-05: channels and columns added or removed, type changes, changed values, fixes, and changes to the index object, the file format and delivery. Newest first. The same changes are listed channel by channel on the changelog by channel.

  • Dates are the day (UTC) a change reached production. A match sits in the revision for the day it was processed, so the revision for a change's date can hold matches from both before and after it. See Changes over time in the spec.
  • About marks a date known only roughly.
  • Releases name the part of the pipeline that changed: the parser, which reads the demo (its version is in header.rushb_version); the converter, which builds the channels (header.ppp_version); the scrubber, which removes personal data; and the publisher, which sends matches to Data Exchange.

2026-10-21​

The converter release that stops writing player_status.is_spotted.

Columns removed

  • player_status.is_spotted: Breaking: no longer written, so code that reads it breaks. It was the game's own "spotted" flag on each player, and it proved unreliable as a sign of who could see whom: checked against real kills, it was never set for one victim who was being shot, and set only a fraction of a second before another kill. Matches published earlier keep the column. Some later manifests still list it, with origin replay-deleted and no data behind it.

2026-10-16​

Parser 6.2.0.

Columns added

  • header.round_time, header.objective: round_time is the round length in seconds that the game itself reports (115 on most maps, 90 in wingman), nullable. It is the value most of the match's rounds started with. It is not always round_time_defuse times 60: on FACEIT and league servers the first round starts under a warmup configuration (two minutes) and the match rules (1.92 minutes, 115 seconds) are applied a few thousand ticks in, so round_time_defuse can hold the warmup value while round_time holds the one the match was played with. objective is what the map asks of the players: bomb, hostage, none or both, nullable. Matches processed before this date have both as null or missing.
  • header.patch_version: Integer: the network protocol of the game build that recorded the demo (14176 to 14181 on the matches checked). build_num is the game build. This is a real value, unlike the old network_protocol column, which was always 1234 and was removed in parser 5.6.0.
  • header.is_pov_recording, header.is_truncated: Booleans. is_pov_recording is true for a recording made on a player's own computer rather than on a server. is_truncated is true when the demo ended early: every table holds what the demo reached, and the match is incomplete. A demo cut in the middle of a message is rejected instead and is not delivered.

2026-10-15​

Parser 6.0.0 and 6.1.0, converter 8.8.0 and 8.8.1.

Channels added

  • convar: Every setting the game server sent, one row per setting each time the server sent it, in the order sent: round, tick, name and value (the server's own string). Valve matchmaking demos send 41 to 47 rows at the start, all at tick 0 except the server's session id, a few dozen ticks later; FACEIT and league servers send 86 to 90 rows, and on FACEIT the match rules (mp_startmoney, cash awards) arrive mid-match after the warmup configuration. The setting in effect at a tick is the last row for that name at or before it. The server's session id (steamworks_sessionid_server) is redacted in the published data. Matches processed before this date don't have the channel.

Columns added

  • header.is_queued_matchmaking, header.queued_matchmaking_mode: Nullable. is_queued_matchmaking is true when Valve's matchmaker queued the lobby (Premier, competitive, wingman) and false on FACEIT and league servers. queued_matchmaking_mode is 1 for Premier and 0 for the others. Both come from the game itself rather than from rank types and row counts.

2026-10-14​

Parser 6.0.0 and the release after it, converter 8.8.0 and the release after it.

Channels added

  • world_object: The map's doors and breakable parts, one row per change: doors opening and closing (spawned, opening, open, closing, closed), the doors built into the map on cs_office and Vertigo moving (moving), and props, vents and Nuke's glass being destroyed (broken), with the player who broke one when there is one (by_player_id). cleanup rows are the map being reset between rounds, not breaks. Matches processed before this date don't have the channel.
  • hostage_vector, hostage_state: Hostages, on hostage maps: hostage_vector has each hostage's position, state, health and carrier every tick, and hostage_state a row each time a hostage changes state (picked up, carried, dropped, rescued, killed) or is hurt, with the player involved. Both are empty on bomb maps. Matches processed before this date don't have the channels.
  • smoke_voxel: What the game sends about each smoke's volume: the cells its fill starts from (kind seed) and, block by block, the map's surfaces around it (kind block, a 4 × 4 × 4 block of 20-unit world cells whose bits mask_lo and mask_hi mark the cells that are part of the map). It is not the smoke itself: filling out from the seeds through free cells for 4 steps gives the shape the game draws, as the spec describes. Rows carry the smoke's grenade_id. Matches from CS2 builds before the voxel data, and smokes already burning when a demo's recording began, have no rows.
  • player_command: One row per user command, what each player's game sent the server for one tick, about a million rows a match. mouse_dx and mouse_dy are the mouse movement in raw counts, forward_move and left_move the movement keys (−1, 0 or 1), weapon_select the weapon switched to (0 when none), and command_number counts each player's commands. The view angles and held buttons are in player_vector and player_inputs. Matches processed before this date, and older CS2 recordings without full user commands, have no rows.
  • player_subtick_input: One row per sub-tick step inside a user command: a button pressed or released (button, the bit as in player_inputs.buttons_mask, and pressed), and when, the moment within the tick as a fraction of it. Steps that carry only analog movement have button 0. Matches processed before this date, and older CS2 recordings without full user commands, have no rows.

Columns added

  • player_vector.is_airborne, player_vector.x_vel_engine, player_vector.y_vel_engine, player_vector.z_vel_engine, player_vector.shots_fired, player_vector.velocity_modifier, player_vector.view_offset_z: The game's own movement and spray state for each player on each tick, all nullable. is_airborne is true when nothing is under the player's feet. x_vel_engine, y_vel_engine and z_vel_engine are the game's velocity at the tick, in units a second. dsdk's x_vel, y_vel and z_vel, which it calculates from positions, describe the move into the tick, half a tick earlier, so the two disagree most after missed ticks, on stairs and in the air. shots_fired counts the shots of the current spray and drops back to 0 soon after the player stops firing. velocity_modifier is the slowdown from being hit, 1 when there is none. view_offset_z is the eye's height above the player's position, about 64 standing and 46 crouched; the eye's height on the map is z_pos + view_offset_z.
  • player_vector.aim_punch_pitch, player_vector.aim_punch_yaw: Floats, in degrees: the recoil kick the player's shots have put on their aim, pitch negative upward. Across a spray they trace the weapon's recoil pattern. The game holds the last shot's value rather than letting it decay, so both are null unless the player fired within the last second (64 ticks), counting the shot's own tick.
  • player_status.progress_bar_start, player_status.progress_bar_duration, player_status.last_injury_time, player_status.is_in_reload, player_status.reserve_ammo, player_status.deploy_tick: The player's and the held weapon's own state, all nullable. progress_bar_start and progress_bar_duration are the defuse progress bar. progress_bar_start is when the player's last defuse began, in seconds on the clock of second, and keeps that value after the defuse ends. progress_bar_duration is the bar's length in seconds while it shows (10, or 5 with a kit) and negative otherwise. CS2 doesn't use the bar for planting. last_injury_time is when the player last took damage, on the clock of second. Both times are null until first set. is_in_reload is the held weapon's reload flag; unlike is_reloading it falls as soon as a reload stops, and about a third of reloads end early in a weapon switch. reserve_ammo is the held weapon's reserve and drops by one magazine per completed reload; null for knives, grenades and equipment. deploy_tick is the tick the held weapon was drawn, joinable on tick.
  • player_info.game_rank_if_win, player_info.game_rank_if_loss, player_info.game_rank_if_tie: Integers, nullable: the rating the game itself predicts the player will hold after the match if their team wins, loses or ties, the "+N / -N" on the Premier scoreboard. Set once at the start of the match, so every round repeats them. 0 outside Premier, as on FACEIT and league demos.
  • header.server_start_tick: Integer, nullable: the game server's tick when the recording's tick 0 was recorded, which differs from match to match. The game counts some times from the server's start; the converter uses this to put player_status.progress_bar_start, last_injury_time and deploy_tick on the recording's clock. Null on a demo with no server tick messages.
  • weapon_fire.inaccuracy_total, weapon_fire.seed, weapon_fire.origin_x, weapon_fire.origin_y, weapon_fire.origin_z, weapon_fire.angle_pitch, weapon_fire.angle_yaw: The server's own record of each shot, on rows where a gun fired a bullet. inaccuracy_total is the inaccuracy the game applied to the shot, including the movement and jumping part that player_vector.inaccuracy (the firing penalty only) leaves out; on about a quarter of shots the total is more than 10% higher. seed is the random seed for the shot's spread, origin_x, origin_y and origin_z are where it was fired from, and angle_pitch and angle_yaw its direction in degrees. Null on grenade, knife and other rows that fire no bullet, and on the rare shot whose record never arrives. Matches parsed before parser 6.0.0 reached production don't have them.
  • weapon_fire.inaccuracy_move, weapon_fire.inaccuracy_air, weapon_fire.aim_punch_pitch, weapon_fire.aim_punch_yaw, weapon_fire.attack_tick, weapon_fire.attack_tick_frac: Sent only by Valve's matchmaking servers, so null on FACEIT and league matches as well as on rows that fire no bullet. inaccuracy_move and inaccuracy_air are the movement and air parts of inaccuracy_total. aim_punch_pitch and aim_punch_yaw are the recoil kick on the view at the shot, in degrees. attack_tick and attack_tick_frac are the tick, and the fraction of it, when the player clicked: almost always 3 ticks before to 1 after the row's tick.
  • player_footstep.player_x_pos, player_footstep.player_y_pos, player_footstep.player_z_pos, player_footstep.player_x_vel, player_footstep.player_y_vel, player_footstep.player_z_vel, player_footstep.player_phi_ang, player_footstep.player_theta_ang, player_footstep.player_weapon_code, player_footstep.player_team_code: The player's position, velocity, view angles, weapon and team at the row's tick, merged from player_vector as on the other player events, so each step, jump and landing has a place. Null where player_vector has no row for that player and tick, about 1 row in 20,000 on a 50-match sample.

Fixes

  • weapon_fire: Two shots by one player on the same tick used to make two identical rows, and the duplicate removal kept only one. The new columns tell them apart, so both are kept. It happens rarely: one match in a 50-match sample gained one row.
  • player_footstep: One row per player, kind and tick. A step the game played twice on one tick, at a jump or a landing, was two rows (about 1 step in 300) and is now one. player_sound still has a row for every time the game played it.
  • player_sound.player_id, player_sound.source: Set on more sounds. A sound whose source reaches the recording later in the same tick, such as a bought weapon or a molotov's fire, gets its player and its source, and an armor purchase goes to the player whose armor rose on that tick. Before, both had no player and source other. On a 50-match sample, other fell from 3.3% of sounds to 0.8%, and 21,250 sounds gained a player; no sound that had a player changed, and no rows were added or removed.

2026-10-12​

Parser 5.7.0 and publisher 4.0.4.

Columns added

  • player_footstep.kind, player_footstep.surface: Strings. kind is step, jump or land. surface is what was stepped or landed on (Concrete, Sand, Wood and so on), and empty for a jump.
  • player_sound.sound, player_sound.source: Strings. sound is the game's name for the sound (Player.PickupWeapon), or its hash as 8 hex digits when the parser's list of names lacks it. source says what made it: player, weapon (one a player carries), grenade, world or other.

Columns removed

  • player_sound.radius, player_sound.duration, player_sound.is_step: Breaking: no longer written, so code that reads them breaks. They came from the player_sound game event, which most demos don't carry.

Types changed

  • player_sound.player_id: Now nullable, and null when no player is known, as for world sounds and about 3% of sounds whose source had not reached the recording when the sound did. The merged position columns are null with it.

Values changed

  • player_footstep: Breaking: Holds every step, jump and landing a player was heard making, about 10,000 rows a match instead of about 560. The rows now come from the sounds the game played. Before, the channel came from a game event that CS2 fires almost only on jumps and landings: of 28,080 such rows on a 50-match sample, 1 was a running step. Code that read every row as a jump or a landing must now filter on the new kind column. A step the game played twice on one tick, at a jump or a landing, is two rows (about 1 step in 300). Matches parsed before this date keep the old channel.
  • player_sound: Holds every sound the game played, about 16,000 rows a match: each player's steps, jumps and landings, pickups and damage heard by others, and the world's and the grenades' sounds. Before, it came from the player_sound game event, which most demos don't carry, so the channel was empty on most matches. Its available flag in the index object is now true on every demo that carries the game's sound messages. A demo with none falls back to the game event, one row per event with sound player_sound. A sound the game played twice on one tick is two rows.

Fixes

  • player_inputs: Has rows again, on every kind of demo: which buttons each player held, a row whenever that changed (6,600 to 82,000 rows a match across a 50-match sample). The CS2 update of 2026-07-09 removed the player property the parser read them from, so the channel, added on 2026-08-03, has been empty on every match recorded since that update, whatever its source; the 2026-08-03 entry's "empty on Valve matchmaking demos" understated it. Parser 5.7.0 (demoinfocs-golang v6) rebuilds the button state from the user commands the demo carries. Matches parsed before this date keep their empty channel.
  • world_item_vector.prev_owner_id: The same on every parse of a demo. Before, on some matches (seen on FACEIT demos) about twenty rows a match could name a different previous owner from one parse to the next. A match is parsed once, so a published match was self-consistent, but processing the same demo again could change this column.

Delivery

  • Each match is delivered as one zip, csds/YYYY/MM/DD/<id>.zip, instead of 43 separate objects. Its entries are the same files under their full keys, stored without recompression, so unzipping it where csds/ sits rebuilds the old layout; dsdk 4.1.0's extract_adx_zips does that, and download_adx_dataset_revision unzips for you. Matches processed after 21:22 UTC are zipped, so the 2026-10-12 revision holds both kinds. See Getting the Data.

2026-10-06​

Converter 8.7.3 and 8.7.4.

Columns removed

  • player_vector.second, player_vector.x_vel, player_vector.y_vel, player_vector.z_vel, player_vector.speed_2d, player_vector.movement_angle, player_vector.movement_angle_diff, player_vector.phi_vel, player_vector.theta_vel, player_vector.ang_vel: Breaking: these ten columns are no longer stored, so code that reads them from player_vector breaks. Each was computed from other columns, and together they were most of the file, which is now under a third of its size. Compute them on load with pureskillgg-csgo-dsdk 3.3.1 or later: add_player_vector_derived_columns(df) adds all ten, from the columns that player_vector_source_columns() lists (tick, round, player_id, the positions and the angles). The index still lists the ten, with origin calculated-deleted.

Types changed

  • weapon_fire.missed_molotov: A boolean, not 0 and 1.
  • molotov_state.was_extinguished_by_smoke, molotov_state.was_extinguished_by_thrown_smoke, molotov_state.was_thrown_into_smoke: Booleans, not 0 and 1. They describe a burn, so they are null, not 0, on every row but inferno_startburn.
  • player_name.name_new, player_name.name_old, other_death.other_type, other_death.weapon_name, player_chat.text, player_disconnect.disconnect_reason, bomb_action.event_type, bomb_defuse.event_type, bomb_state.event_type, header.bomb_time, header.ppp_version, header.rushb_version: Written as the column's own type when a match has no value for it or the channel is empty. Before, such a column was written with Parquet's null type, so its type changed from match to match.

Index object

  • Every column's entry gives the type it is written as and whether it can be null, and every merged and calculated column is declared nullable. Before, most columns declared a narrower type than the one written (int32, float32) or one with no width (int), and merged and calculated columns declared neither.

2026-10-05​

Parser 5.4.1 to 5.6.0 and converter 8.5.4 to 8.7.1.

Columns added

  • grenade_vector.grenade_id, grenade_bounce.grenade_id, grenade_state.grenade_id: One id per thrown grenade for the whole match (int64). Join the three grenade channels on it. From parser 5.6.0 the grenades are numbered 1, 2, 3 in the order the parser first meets them, so a match parsed again gets the same ids; parser 5.5.0, in production for part of 2026-10-05, drew them at random on each parse. entity_id is reused within a match, so grouping on it can merge different grenades. It can be null on grenade_state.
  • grenade_vector.grenade_weapon_code: The grenade's Valve item id (int64): 43 flashbang, 44 HE grenade, 45 smoke grenade, 46 molotov, 47 decoy and 48 incendiary. These are the same codes every other *_weapon_code column uses.
  • header.rank_type: The ladder the match was played on (int64), the rank type its players carry: 7 Wingman, 11 Premier, 12 Competitive; 0 when nobody is ranked; -1 off Valve's ladders, as on FACEIT. Null when the demo has no player information.
  • molotov_state.extinguisher_not_found: True where the fire ended early but no smoke was near enough to credit (bool). Before, extinguisher_id, extinguisher_id_fixed, smoke_entity_id and smoke_entity_id_fixed held -2 there.

Columns removed

  • grenade_vector.grenade_type_code: It held the parser library's own grenade numbering, 501 to 506, which collides with the item ids of knives. Read grenade_weapon_code instead.
  • header.network_protocol: It was 1234 in every match, a fixed value rather than the demo's.
  • item_equip.def_index: It was 0 on every row, because CS2's equip event carries no item id. item_pickup.def_index carries it.

Types changed

  • rank_update.rank_change: An integer (int64), like rank_old and rank_new. Before, a double holding whole numbers.
  • bomb_action.player_id_fixed, bullet_damage.player_id_fixed, grenade_bounce.player_id_fixed, grenade_state.player_id_fixed, grenade_vector.player_id_fixed, item_dropped.player_id_fixed, item_equip.player_id_fixed, item_pickup.player_id_fixed, item_refund.player_id_fixed, molotov_fire.player_id_fixed, player_blind.player_id_fixed, player_chat.player_id_fixed, player_connect.player_id_fixed, player_death.player_id_fixed, player_death.attacker_id_fixed, player_death.assister_id_fixed, player_disconnect.player_id_fixed, player_footstep.player_id_fixed, player_hurt.player_id_fixed, player_hurt.attacker_id_fixed, player_inputs.player_id_fixed, player_sound.player_id_fixed, player_spawn.player_id_fixed, player_status.player_id_fixed, player_vector.player_id_fixed, rank_update.player_id_fixed, team_change.player_id_fixed, weapon_action.player_id_fixed, weapon_fire.player_id_fixed: An integer in every match, with nulls where no player matched. Before, a match in which any row found no player wrote the column as a double, so its type changed from match to match.
  • player_vector, player_status: Written compactly, with the same values. Each integer column is the smallest signed type that holds it, int8, int16 or int32, where it was int64. The floats read from the demo are float32, where they were double. place_name is dictionary-encoded. The spec's tables give every column's type. pandas reads nullable Int8, Int16 or Int32 where it read Int64, and a category for place_name. Element-wise arithmetic stays narrow and can wrap (with int16 money, money * 5 gives 14,464 for 16,000), so cast to a wider type first. How to read old and new files together is in The player tables.

Values changed

  • player_vector.x_vel, player_vector.y_vel, player_vector.z_vel, player_vector.theta_vel, player_vector.phi_vel, player_vector.ang_vel, player_vector.speed_2d, player_vector.movement_angle: Computed per player per round, so a player's first sample in each round is 0. Before, positions and view angles were differenced across rounds, and a respawn read as a jump, as much as 365,000 units a second in our test matches. A velocity component over 3,500 within a round, the game's sv_maxvelocity, is taken as a teleport and reads 0. speed_2d, movement_angle and ang_vel follow from these.
  • bomb_action, bomb_defuse, bomb_state, bot_takeover, bullet_damage, grenade_state, item_dropped, item_equip, item_pickup, item_refund, molotov_state, other_death, player_blind, player_chat, player_death, player_hurt, player_inputs, player_sound, player_spawn, rank_update, team_change, weapon_action, weapon_fire: The *_x_vel, *_y_vel and *_z_vel columns merged from player_vector take its per-round velocities, so they no longer jump on a respawn.
  • player_vector.movement_angle_diff: The angle between where the player looks (theta_ang) and where they move, in -180 to 180: 0 moving the way they look, ±90 sideways, ±180 backwards, and null while standing still. Before, it was theta_ang minus movement_angle with 360 added once to a negative result, so it ran from -180 to 360, the same angle could read 350 or -10, and standing still read -1.
  • player_hurt.attacker_id, player_hurt.attacker_id_fixed, player_death.attacker_id, player_death.attacker_id_fixed, player_death.assister_id, player_death.assister_id_fixed, other_death.attacker_id, other_death.attacker_id_fixed: Null where there is no attacker or assister, not 65535. Declared nullable.
  • molotov_state.player_id, molotov_state.player_id_fixed, molotov_state.tick_throw, molotov_state.burn_duration: Null, not -1, where no thrower was found: every inferno_expire row, and a round whose throws and fires don't pair up. burn_duration is null on inferno_expire rows; every start row still has one. Declared nullable.
  • molotov_state.extinguisher_id, molotov_state.extinguisher_id_fixed, molotov_state.smoke_entity_id, molotov_state.smoke_entity_id_fixed: Null on inferno_expire rows and where nobody put the fire out, not -1 or 0; the -2 moved to extinguisher_not_found. A real player 0 who put a fire out keeps extinguisher_id 0, where before it read the same as nobody. Declared nullable.
  • grenade_state.tick_throw: Null where no throw was matched, not 0, -2 or -3. Declared nullable.
  • tick.second_since_previous_phase: Null before the match's first phase, not -1. Declared nullable.
  • player_hurt.weapon_name: Null for damage with no weapon, not an empty string.

Fixes

  • bomb_action.player_id: Set on bomb_pickup rows. It was null on all of them, so those rows joined to no player and carried no player position.
  • bullet_damage.num_penetrations: Counts the surfaces the bullet went through. It was 0 on every row.
  • round_state.event_type: The freezetime_ended_inferred row, the first tick a player moves in a round, no longer fires early on a late player's respawn. In 8 of the 927 rounds of our 50 test matches it came 43 to 3,520 ticks before round_freeze_end; it now lands one tick after it.
  • player_vector.current_ammo: The number of bullets in the magazine, with 0 for an empty one. Before, every value was one low and an empty magazine read 4294967295. A replay parsed before this change and converted again after it holds -1 for an empty magazine, with the other values still one low.
  • player_status.inv_flashbang, player_status.equipment_value_calc: inv_flashbang counts the flashbangs held, up to 2, and drops to 0 at the throw of the last one. Before, it read 1 for two flashbangs and could stay at 1 for about 0.7 seconds after the throw. equipment_value_calc, which counts them, follows.
  • player_spawn.round: The round the player spawned into. Before, many rows carried the round before it (7,913 in our 50 test matches).
  • player_info, player_personal: Every player who spawns in a round that reaches freeze end has a row for that round. Before, a player missing from the round's roster, such as one who joined or reconnected late, had none, so their rows in that round joined to no player.

File format

  • player_vector, player_status: Rows are sorted by player_id, then tick, where they were in tick order; sort by tick, then player_id, for tick order. The floats are stored with Parquet's byte stream split encoding, which fastparquet can't read. pyarrow (pandas' default engine), polars and DuckDB read both the old and the new files.

2026-10-03​

Converter 8.5.3.

Values changed

  • header.t_starters_avg_rank, header.ct_starters_avg_rank, header.t_starters_avg_wins, header.ct_starters_avg_wins: Averaged over the first half's players only, rounds up to max_rounds / 2 (12, or 8 in Wingman). Before, CS:GO's 15-round half was used, which mixed in the other team's players, so both teams often got the same average. A team with no players is now null instead of -1.
  • header.t_starters_score_final, header.ct_starters_score_final: Each starting team's own final score, followed through the half-time and overtime side swaps. Before, it was the larger of the side scores, which could give a team the other team's score.
  • header.tick_save_rate: Now 64, since every tick is kept. It was a fixed 21 that did not describe the data.

Fixes

  • grenade_state.entity_id_fixed: A decoy with no decoy_started event keeps its -2 instead of being overwritten.
  • player_info.rank_platform, player_info.elo_platform: On FACEIT matches, each player's FACEIT level and Elo now belong to the right player. Before, they were matched by row position and could be another player's.

2026-09-07​

Parser 5.3.7 to 5.4.0, converter 8.3.1 to 8.5.0, publisher 4.0.3.

Columns added

  • player_info.player_controller_id: The player's controller id, as already in player_status and player_personal. Tells apart two players who share a player_id in one round.
  • header.freeze_time, header.round_time_defuse, header.bomb_time, header.max_rounds: The server's own settings for freeze time (seconds), round time (minutes), bomb timer (seconds) and maximum rounds. Null when the demo does not record the setting; bomb_time has been null in every match checked.

Fixes

  • player_info: A match in which two different players share a round and player_id_fixed now fails processing and is not published.

Index object

  • tick: The tick channel's category is now telemetry (was single_event).

Delivery

  • The daily revision now closes reliably. Before, when a close failed, that day's matches were published in the next revision, so one revision could hold more than one day.

2026-08-22​

Parser 5.3.4 to 5.3.6, converter 8.2.3 and 8.3.0.

Columns added

  • player_status.is_reloading: Whether the player is reloading (bool). Replaces reload_visually_complete.
  • header.is_gotv_recording: True when the match was recorded with GOTV transmission restricted, as on tournament servers. Such matches lack the per-player channels GOTV restricts, such as player_footstep and item_equip.
  • header.is_wingman: True for a Wingman match.
  • header.max_unique_players: Distinct human players over the whole match. Above the lobby size when someone was substituted or came back on another account.
  • molotov_state.smoke_entity_id_fixed: Joinable id of the smoke that put the fire out, matching its grenade_state.entity_id_fixed. It is 0 when the fire was not put out and -2 when no smoke was found. smoke_entity_id is the raw id, which the game reuses within a match.
  • molotov_state.fraction_extinguished: Share of the fire's flames, 0 to 1, that a smoke put out.
  • molotov_state.was_thrown_into_smoke: 1 when the molotov or incendiary landed inside a smoke that was already up.
  • molotov_fire.entity_id_fixed: The molotov_state.entity_id_fixed of the fire this flame belongs to, or -1 if unmatched. Join on this, not on entity_id, which the game reuses.

Columns removed

  • player_status.reload_visually_complete: Null on every row since 2025-08-21. Use is_reloading.
  • player_vector.second_diff, tick.next_valid_tick, tick.next_valid_tick_round: Internal working columns that were written but never listed in the index.
  • header.__index_level_0__, player_vector.__index_level_0__: A leftover row-number column, never listed in the index, is gone from the last two channels that still had it.

Types changed

  • bomb_defuse.has_kit, other_death.is_attacker_blind, other_death.is_noscope, other_death.is_through_smoke, player_death.is_attacker_blind, player_death.is_flash_assist, player_death.is_noscope, player_death.is_through_smoke, player_chat.is_chat_all, player_status.burst_mode, player_status.is_silenced: Was int64 holding 0 or 1, now bool, as the index always declared. Still nullable.

Values changed

  • molotov_state.was_extinguished_by_smoke: No longer set for every incendiary that simply burned out. Reads 0 rather than -1 when the fire was not put out. Where molotov_fire is present, a fire counts as put out when any of its flames was, so partial extinguishes count. Smokes are taken to last 17.47 seconds (was 17).
  • molotov_state.extinguisher_id, molotov_state.extinguisher_id_fixed, molotov_state.smoke_entity_id: Only a smoke within 250 units of the fire is credited; otherwise -2. Before, smokes far from the fire were credited (the median was over 1,000 units away).
  • molotov_state.was_extinguished_by_thrown_smoke: Written in every match, as 0 or 1. Before, it was listed in the index but written only in some matches, and was never 1.

Index object

  • header, player_vector, round_state: The index stops listing 62 columns that were never written. They are the header's protocol, playback_time, playback_ticks, playback_frames, signon_length and second, and the player_tick and player_player_id entries (and their attacker_ and assister_ forms) in 21 event channels. Duplicate entries are gone too: player_vector's velocities, round_state.second and several header columns.

2026-08-03​

Parser 3.0.0 to 5.3.1, converter 8.0.0 to 8.2.2, scrubber 3.1.0.

Channels added

  • bullet_damage: Per-bullet detail for each player_hurt: distance, damage direction, penetrations, and no-scope and in-air flags. Valve matchmaking demos only.
  • grenade_bounce: Every surface a thrown grenade hit.
  • grenade_vector: The flight path of each thrown grenade, a row whenever its position changed.
  • item_dropped: A weapon leaving a player's hands, with the reason (death, replaced or manual).
  • item_refund: Buy refunds.
  • molotov_fire: Each flame of each molotov or incendiary fire, when it ignites and when it goes out.
  • player_chat: In-game chat. text reads redacted in the published data.
  • player_connect: Player and bot connects, warm-up included.
  • player_inputs: Which buttons each player held, a row whenever that changed. Empty on Valve matchmaking demos, which do not carry it.
  • player_sound: Sounds a player made and how far they carried. Empty on Valve matchmaking demos, which do not carry it.
  • rank_update: Rank changes at the end of the match, official matchmaking only. win_count above 2500 is published as 2501, as player_info.wins is.
  • score_update: Each change in a team's score.
  • team_change: Players switching teams, including the silent half-time swap.
  • world_item_vector: Positions of items lying on the ground, a row whenever one moved.

Channels removed

  • item_remove: It never had rows in CS2 data; it was a placeholder kept from CS:GO.
  • player_action: Empty since CS2 demos stopped carrying these events (see 2026-01-22).

Columns added

  • round_mvp.mvp_count: The player's running MVP total. round_mvp has rows again, one whenever a player's MVP count rises; it had been empty on most matches since early 2024.
  • bot_takeover.is_controlling: True when a player takes over a bot, false when they give it back. bot_takeover has rows for the first time in CS2 data.

Columns removed

  • round_mvp.mvp_reason_code, round_mvp.music_kit_mvps: No CS2 source.
  • round_end.legacy_code: No CS2 source.
  • round_start.time_limit, round_start.frag_limit, round_start.objective: Always 0, 0 and empty in CS2 data.

Values changed

  • round_end.win_reason_code, round_end.win_reason_message, round_end.player_count: Real round ends again. win_reason_code and win_reason_message come from the game, and player_count is the number of players on the team at the end of the round, bots included. Since 2024-02-12 the channel had been rebuilt from round_state with fixed values.
  • round_start: Real round starts again, one row per round, round 1 included. The rebuilt channel used since 2024-02-12 was one row short.
  • round_state: The round_start and round_end event rows are back (missing since 2024-02-12). Each round now has two round_start rows: the game's own, and the one the converter adds after round_poststart (since 2024-08-05).
  • player_name: Records name changes. It had no rows in any CS2 data before.

Fixes

  • round_state.t_score, round_state.ct_score: On round_end rows, the final round is no longer counted twice in 24-round matches (a CS:GO 30-round rule had been applied).

Index object

  • bot_takeover, bullet_damage, player_inputs, player_sound, rank_update, round_mvp: A channel entry can carry available. False means the demo did not carry that data, so an empty channel means "not recorded" rather than "nothing happened". It is set on bot_takeover, bullet_damage, player_inputs, player_sound, rank_update and round_mvp.

2026-07-08​

Converter 7.2.0.

File format

  • Parquet files are compressed with ZSTD instead of GZIP. Values are unchanged; your Parquet reader needs ZSTD support.

2026-07-07​

Parser 2.10.0 to 2.12.2, converter 7.0.0 to 7.1.1.

Columns added

  • player_info.rank_type: Which ladder rank is on: 11 Premier (CS Rating), 12 Competitive (skill group 0 to 18), 7 Wingman; 0 or -1 when unknown. Every player in a match has the same value.
  • player_info.elo_platform: The player's FACEIT Elo, on FACEIT matches only. The column is absent from other matches.

Columns removed

  • The leftover row-number column __index_level_0__, never listed in the index, is no longer written in most channels. header and player_vector kept it until 2026-08-22.

Values changed

  • tick.previous_phase, tick.second_since_previous_phase: When several phase events share a tick, the phase now follows a fixed order instead of whichever event sorted last.

File format

  • Files are written by a newer Parquet library, in format version 2.6 instead of 1.0.

2026-06-07​

Converter 6.11.4.

Fixes

  • player_personal.player_id, player_status.player_id, player_vector.player_id: When a round records the same player_id for two players, the player's controller id is used instead. A match that would still have duplicate tick and player_id_fixed rows fails processing and is not published.

2026-06-05​

Converter 6.11.3.

Values changed

  • header.providence, player_info.rank_raw, player_info.rank_platform: A match without the platform's metadata file is now published, without these columns. Before, it failed processing and was not published.

Fixes

  • grenade_state.entity_id_fixed: -2 for a decoy with more than one start event. Such matches used to fail processing and were not published.
  • player_vector.team_code: Matches with a bot takeover are published; they used to fail processing.

2026-04-29​

Parser 2.8.0 and 2.9.0.

Columns removed

  • player_vector.x_aimpunch, player_vector.y_aimpunch, player_vector.aim_punch_angle_vel_x, player_vector.aim_punch_angle_vel_y, player_vector.aim_punch_angle_vel_z: The aim-punch columns are no longer written.

2026-01-25​

Parser 2.7.1.

Values changed

  • header.network_protocol: Always 1234; the demo no longer provides a meaningful value.

About 2026-01-22​

No release; a change in what cs2 demos carry.

Change in CS2 demos

  • player_action: CS2 demos stopped carrying these events around this date; most matches from here on have no rows. The channel was removed on 2026-08-03.

2025-08-21​

Parser 2.6.0.

Values changed

  • player_status.reload_visually_complete: No longer filled, so null on every row. Removed on 2026-08-22 in favour of is_reloading.

2024-08-05​

Converter 6.10.0 to 6.10.7.

Columns added

  • tick.next_valid_tick, tick.next_valid_tick_round: Internal working columns, written but not listed in the index. Removed on 2026-08-22.

Values changed

  • player_id, attacker_id and assister_id, and with them every *_id_fixed and merged position column, are remapped round by round to the game's controller numbering, so events join to the right player.
  • round_state: A round_start row is added for each round, one tick after round_poststart.
  • round_start: When the demo has no round starts, the channel is rebuilt from those new round_state rows instead of from round_prestart.

Fixes

  • Event rows whose player cannot be identified are kept, without the player's position columns.

2024-07-28​

Converter 6.9.2 to 6.9.4.

Fixes

  • The id remapping is skipped when a match's ids are already small, as on FACEIT.
  • round_state: freezetime_ended_inferred ignores the first 5 ticks of a round, where the spawn teleport looked like movement.

2024-07-07​

Converter 6.9.1.

Columns added

  • header.unique_steamids: Number of distinct players in player_personal.

Values changed

  • round_state.event_type: New event freezetime_ended_inferred: one row per round, at the first tick any player moves.

Fixes

  • header.t_starters_score_final, header.ct_starters_score_final: Missing scores no longer spoil the result.

2024-06-06​

Parser 2.4.2.

Fixes

  • Demos recorded after the CS2 update of 2024-05-29 parse again. The parser library update also fixes player data that could be corrupted or missing.

2024-04-28​

Parser 2.4.0 and 2.4.1.

Fixes

  • Demos recorded after the CS2 update of 2024-04-25 parse again.

About 2024-03-01​

No release; a change in what cs2 demos carry.

Change in CS2 demos

  • round_mvp: From February 2024 a growing share of matches have no rows; by July 2024 none of the matches checked do. The channel has rows again from 2026-08-03.

About 2024-02-24​

Publisher 1.2.4 and 1.3.0.

Delivery

  • Publishing fixes after AWS changed Data Exchange permissions. Matches from the days before may have been published late.

2024-02-12​

Parser 2.2.0, converter 6.7.0 and 6.8.0.

Values changed

  • round_end.win_reason_code, round_end.win_reason_message, round_end.legacy_code, round_end.player_count: CS2 demos stopped carrying round-end events. The channel is rebuilt from round_state: win_reason_code is 8 (CT win) or 9 (T win), win_reason_message is the generic team-win message, legacy_code is 0 and player_count is 20.
  • round_start: CS2 demos stopped carrying round-start events. The channel is rebuilt from round_state and has one row fewer than the rounds played.
  • round_state: No more round_start and round_end event rows: CS2 demos stopped carrying them.

Fixes

  • player_info.team_code: The parser library update fixes players who could be given the wrong team, and disconnected players who were still treated as playing.

Delivery

  • No matches are dated 2024-02-08 to 2024-02-11: the CS2 update of 2024-02-06 broke demo parsing until this release.

2023-12-09​

Parser 2.1.0, converter 6.6.0.

Columns added

  • player_status.player_controller_id, player_personal.player_controller_id: The player's controller id, stable for a player slot. Joins player_status rows to player_personal.

Values changed

  • player_blind: Keeps only flashes on players alive at that tick. CS2 also reports flashes on dead and spectating players.

2023-11-23​

Converter 6.5.6.

Values changed

  • In some matches player ids came out about 65,000 too high; they are corrected, so events join to the right player.

2023-11-05​

Parser 2.0.0, converter 6.5.5.

CS2 data begins

  • The first CS2 matches are published, with 30 channels, described in the archived spec. item_remove, player_name and bot_takeover have no rows. player_vector.second_diff and a __index_level_0__ column in every channel are written but not listed in the index, and header.tick_save_rate is a fixed 21.