Three independent changes bundled here — performance, maintainability, and
deployment DX.
1. Hot-path Allocations (ad_tag caching):
Every middle-relay session was doing a hex::decode of the user's ad_tag (plus a fallback to the global one) on every single connection. That's two string allocations and two hex decodes per session for data that only changes on config reload. Moved the decoding into ProxyConfig load/hot-reload into a precomputed runtime_ad_tags cache (same #[serde(skip)] pattern as runtime_user_auth). The session path now does a single O(1) lookup via config.effective_ad_tag(&user). Falls back to on-demand decode if the cache isn't built (e.g. in tests), so existing behavior is preserved. This eliminates two useless allocations and hex decodes per connection.
2. Metrics Monolith Split(structural refactor):
Broke up the massive 4,200-line `metrics.rs` into a proper `metrics/` directory module (`mod.rs`, `tls_front.rs`, `tests.rs`). I isolated the giant `render_metrics` function into its own `render.rs` file. This was a purely structural move—no locks or function bodies were changed, keeping things build-safe while making the crate significantly easier to navigate. The /metrics endpoint output is byte-identical.
3. GitOps / Dokploy Auto-Deployment Support (deployment DX):
Added `figment` to `Cargo.toml` to seamlessly merge TOML config files with Environment Variables. You can now configure the proxy entirely using `TELEMT_` prefixed ENV variables (e.g., `TELEMT_GENERAL__PORT=443`) without needing to manually mount a physical `config.toml` via Docker volumes. The Dockerfile and compose setups now gracefully handle missing config files by generating an empty placeholder, making stateless auto-deployments on platforms like Dokploy frictionless.
client_mss (e.g. "tspu", MSS=92) fragments the whole connection to evade
DPI on the ServerHello, but it also fragments bulk payload, multiplying
outgoing packets-per-second ~10x. On hosts whose abuse detection counts
pps (not bandwidth) this trips packet-flood limits.
Add an optional [server].client_mss_bulk: keep the low client_mss for the
handshake (ServerHello stays fragmented => DPI bypass intact), then raise
the client socket MSS to client_mss_bulk once the connection enters the
post-handshake (bulk transfer) phase, so bulk data uses normal-size
segments and pps drops back to normal. Same preset/int grammar as
client_mss. Opt-in: when unset, the handshake MSS is kept for the whole
connection (unchanged behavior).
Linux-only (setsockopt TCP_MAXSEG via raw fd, mirroring TCP_USER_TIMEOUT);
no-op on other unix. Documented in CONFIG_PARAMS.{en,ru}.
- bound API and metrics connection handling
- default metrics listener to localhost
- reject untrusted PROXY protocol peers before parsing headers
- cap API request body size and PROXY v2 payload allocation
- validate route usernames and TLS domains consistently
- Relocate tg_connect from [timeouts] to [general] with validation and docs updates.
- Apply rustfmt to per-attempt upstream connect timeout expression in upstream.rs.
- Pass tg_connect_timeout_secs in all UpstreamManager::new test call sites.
- Wire hot reload and runtime snapshot to general.tg_connect.
Wire config.timeouts.tg_connect into UpstreamManager; per-attempt timeout uses
the same .max(1) pattern as connect_budget_ms.
Reject timeouts.tg_connect = 0 at config load (consistent with
general.upstream_connect_budget_ms and related checks). Default when the key
is omitted remains default_connect_timeout() via serde.
Fixestelemt/telemt#439
ali_rajabpour