If players survive a whole map but get dropped the moment it changes — "Connection timed out", stuck on the loading screen, or a mass disconnect at changelevel — the problem is almost always the resource transfer that happens between maps. The new map's custom models, sounds and sprites are being pushed to clients over the slow in-game UDP channel, the transfer runs past the timeout, and the client gives up. Less often it is a precache overflow that takes the whole server down. Both are fixable.
1. Set up a fast download (fastdl) server
By default, custom content is sent to clients over the game socket — a trickle that easily exceeds the connection timeout on a content-heavy map. A fastdl (HTTP) mirror moves that transfer to a real web server. In server.cfg:
sv_downloadurl "http://your-cdn.example.com/cstrike/" sv_allowdownload 1 sv_allowupload 1
Mirror your cstrike/ content tree (maps, models, sounds, sprites, gfx) under that URL, preserving the folder structure exactly. Clients then fetch resources over HTTP in seconds instead of timing out on UDP. This is the single most effective fix for map-change timeouts on custom-content servers.
2. Give the transfer more room
Even with fastdl, raise the timeout and the per-client upload allowance so a client on a slower link is not cut off mid-transfer:
sv_timeout 120 sv_max_upload 5 sv_allowdownload 1
sv_timeout is how long the server waits before declaring a silent client dead — and a client busy downloading a large map looks silent to the game socket. Give it headroom. Do not set it absurdly high on a public server, though, or genuinely dropped clients linger in slots and hold them against your slot count for two minutes each. The client's own cl_dlmax and rate settings matter too, but those live on the player's machine and you cannot force them from the server; a working fastdl mirror is what removes the dependency on them.
Keep your network rates sane in server.cfg as well, because a client that renegotiates rate at map load on a starved connection is more likely to stall out during the resource push:
sv_maxrate 25000 sv_minrate 15000 sv_maxupdaterate 101 sv_minupdaterate 30
3. Rule out a precache overflow
The other cause is the server exceeding an engine table when it loads the new map. GoldSrc caps precached models and sounds, and a plugin (or a heavy map) that precaches too many resources crashes the server at map load with lines like:
PF_precache_model: Model overflow SZ_GetSpace: overflow nummodels == MAX_MODELS
If your log shows one of these at the exact moment of the change, the fix is to reduce precaches, not to tune networking. Audit resource-heavy plugins (custom models, skins, sound packs) and trim what each precaches. A plugin that precaches a model per weapon skin across dozens of skins is the usual offender. Bisect with the crashing-plugin method.
4. Consistency kicks look like timeouts
If the server enforces file consistency and a client's copy of a model differs, the client is dropped at map load with a consistency error that reads to players as "it kicks me on map change". Check for Server is enforcing consistency for this file in the client console. Either ship the exact files via fastdl (step 1) so clients match, or relax the consistency enforcement for the offending files.
Common errors
- Only content-heavy maps drop players — no fastdl; the UDP download times out. Step 1.
- Every map change crashes the whole server, not just one client — precache overflow, step 3. Check the log for
Model overflow/SZ_GetSpace. - fastdl set but clients still download in-game — the URL structure does not match your
cstrike/tree, so the HTTP fetch 404s and the client falls back to UDP. Fix the mirror layout and check for 404s in the web server log. - Players on fast connections are fine, slow ones drop — classic timeout-too-tight; raise
sv_timeoutand make sure fastdl is actually serving.
Verification
Connect a client on an ordinary connection and force several map changes across your heaviest custom maps:
changelevel <heavy_custom_map>
The client should reach the new map without timing out, and with fastdl working you will see it pull files over HTTP (fast) rather than the slow in-game bar. On the server, watch Log/errors-*.log for precache overflow lines during the change — there should be none. A client that rides through five consecutive changes on your worst maps, with no overflow in the log, means both the transfer and the precache load are under control.
