Advance forge NBC worker and Ethereum full nodes
This commit is contained in:
every.channel
parent
7d84510eac
commit
3402f7dab2
17 changed files with 3066 additions and 414 deletions
|
|
@ -0,0 +1,15 @@
|
|||
# ECP-0101: Decode Archived Object Frames Before HLS Replay
|
||||
|
||||
## Why
|
||||
|
||||
`wt-archive` stores raw MoQ object frames in CAS. `wt-archive-serve` was replaying those framed bytes directly as `init.mp4` and `.m4s`, which yields HLS playlists that look valid but cannot be decoded by players.
|
||||
|
||||
## Decision
|
||||
|
||||
Decode archived object frames back into their media payload bytes before serving replay responses. If archived bytes are already raw payloads, pass them through unchanged.
|
||||
|
||||
## Consequences
|
||||
|
||||
- Local and forge replay can hand archived CMAF data to normal HLS consumers.
|
||||
- Existing archives remain usable without rewriting CAS objects.
|
||||
- Replay failures now point at genuinely bad archive contents rather than envelope bytes leaking over HTTP.
|
||||
|
|
@ -0,0 +1,35 @@
|
|||
# ECP-0102: Linux Widevine NBC Worker on `ecp-forge`
|
||||
|
||||
## Why
|
||||
|
||||
NBC live playback currently exists only in the desktop app:
|
||||
|
||||
- macOS native `WKWebView` for in-app auth and playback capture
|
||||
- local Chrome fallback for browser-driven playback capture
|
||||
|
||||
`ecp-forge` cannot originate an NBC live stream today because `ec-node` has no NBC source worker, no Adobe auth bootstrap flow, and no Linux browser runtime with Widevine.
|
||||
|
||||
## Decision
|
||||
|
||||
Build a forge-capable NBC source worker around Linux Google Chrome plus a persistent authenticated profile.
|
||||
|
||||
1. Use Google Chrome on `x86_64` Linux, not bare Chromium, as the browser runtime for NBC on forge.
|
||||
2. Run Chrome in a virtual display session on `ecp-forge` so DRM playback remains in a normal browser path even when no physical display is attached.
|
||||
Keep the forge worker on a dedicated display number and disable GPU acceleration by default under Xvfb.
|
||||
3. Persist the NBC/Adobe browser profile on forge so auth warmup is amortized across runs.
|
||||
4. Add an explicit bootstrap path that surfaces interactive auth only when the stored session is cold or expired.
|
||||
5. Launch Chrome as an external process and attach over DevTools rather than relying on the crate-managed Chrome launcher path.
|
||||
6. Publish the resulting captured stream through the existing `ec-node` relay path so archive, replay, and downstream nodes stay unchanged.
|
||||
7. Expose the forge operator surface as `ec-node nbc-bootstrap` and `ec-node nbc-wt-publish`, with a persistent virtual display on `ecp-forge`.
|
||||
|
||||
## Consequences
|
||||
|
||||
- NBC live origin becomes possible from forge without depending on a user’s macOS desktop session.
|
||||
- We keep the existing every.channel transport/archive model and only replace the source worker.
|
||||
- The first forge implementation should be browser-frame capture first; audio and full unattended auth renewal can follow once the Linux worker is stable.
|
||||
|
||||
## Rejected Alternatives
|
||||
|
||||
- Reuse the macOS `WKWebView` path on forge: impossible on Linux.
|
||||
- Depend on bare Chromium only: rejected because Widevine support is the central requirement.
|
||||
- Require true headless-only DRM playback from the start: rejected because the safer first implementation is a normal Chrome session inside a virtual display.
|
||||
|
|
@ -0,0 +1,24 @@
|
|||
# ECP-0103: Mullvad Philadelphia Egress for Forge NBC Philadelphia
|
||||
|
||||
## Why
|
||||
|
||||
The forge-side NBC worker is currently dependent on a reverse-tunneled proxy for US egress.
|
||||
That is enough to prove the geo-boundary, but it is the wrong long-term operator shape for `NBC Sports Philadelphia`.
|
||||
|
||||
## Decision
|
||||
|
||||
1. Enable the Mullvad daemon on `ecp-forge`.
|
||||
2. Keep the Mullvad account number out of committed Nix configuration; log in operationally from founder-provided material.
|
||||
3. Use a Philadelphia Mullvad relay for `NBC Sports Philadelphia` work on forge.
|
||||
4. Start the forge NBC publish worker after the Mullvad daemon is available.
|
||||
|
||||
## Consequences
|
||||
|
||||
- Forge NBC egress becomes self-contained instead of depending on a local reverse proxy.
|
||||
- The account credential stays operational-only rather than being copied into repo config.
|
||||
- Relay choice remains runtime-controlled, so it can be swapped if a specific Philadelphia host degrades.
|
||||
|
||||
## Rejected Alternatives
|
||||
|
||||
- Keep relying on the reverse-tunneled local proxy: rejected because it couples forge origin to a founder workstation.
|
||||
- Commit the Mullvad account number into NixOS config: rejected because it expands secret exposure for no benefit.
|
||||
|
|
@ -0,0 +1,45 @@
|
|||
# ECP-0104: `ecp-forge` dual Ethereum full nodes on dedicated NVMe ZFS
|
||||
|
||||
## Why
|
||||
|
||||
`every.channel` now has OP Stack and Ethereum contract rails on `ecp-forge`, but it still depends on
|
||||
public L1 RPC and beacon providers for Ethereum itself.
|
||||
|
||||
The forge host has enough raw capacity in `tank`, but `tank` is a large HDD `raidz3` pool that is
|
||||
well suited for archive bytes, not for the random-read/write pattern of concurrent Ethereum
|
||||
execution and consensus sync. The same host also has a free 7 TB NVMe device, which is the right
|
||||
media class for a self-hosted node.
|
||||
|
||||
## Decision
|
||||
|
||||
1. Create a dedicated single-device ZFS pool for Ethereum on the free NVMe in `ecp-forge`.
|
||||
2. Run two full-sync Ethereum node pairs on that pool:
|
||||
- Ethereum mainnet
|
||||
- Ethereum Sepolia
|
||||
3. Use Reth for execution and Lighthouse for consensus.
|
||||
4. Keep raw execution and beacon RPC local-only on `ecp-forge` for the first tranche.
|
||||
5. Publish sync and finality status on `https://eth.every.channel`.
|
||||
6. Because this tranche is explicitly full-sync from genesis, run Lighthouse with
|
||||
`--allow-insecure-genesis-sync` instead of checkpoint bootstrap.
|
||||
|
||||
## Consequences
|
||||
|
||||
- `every.channel` gets repo-owned L1 execution and consensus rails for both mainnet and Sepolia.
|
||||
- Ethereum state no longer competes with `tank` archive I/O.
|
||||
- The first public surface is intentionally conservative: health and sync visibility, not
|
||||
unauthenticated public JSON-RPC.
|
||||
- The dedicated Ethereum pool is single-device NVMe, so it optimizes for node performance rather
|
||||
than storage redundancy.
|
||||
- Consensus sync starts from genesis without checkpoint assistance, which is slower but matches the
|
||||
requested full-sync posture.
|
||||
|
||||
## Rejected Alternatives
|
||||
|
||||
- Put Ethereum state on `tank`: rejected because the HDD `raidz3` pool is the wrong latency profile
|
||||
for concurrent execution and consensus sync.
|
||||
- Keep using only public upstream Ethereum providers: rejected because the OP Stack and settlement
|
||||
rails should not depend on third-party RPC availability.
|
||||
- Expose raw JSON-RPC on `eth.every.channel` immediately: rejected for the first tranche because it
|
||||
creates an unauthenticated public abuse surface before auth and rate policy exists.
|
||||
- Use checkpoint sync for Lighthouse: rejected for this tranche because the requested posture is
|
||||
full sync from genesis on both networks.
|
||||
Loading…
Add table
Add a link
Reference in a new issue