CDNs grew up on short HTTP transactions; modern applications hold sockets open for hours and stream binary frames both ways. Running WebSockets and gRPC through an edge is entirely doable in 2026, and entirely full of details that separate providers who support it from providers who tolerate it.
WebSockets: proxying, not caching
A WebSocket starts as an HTTP Upgrade and becomes a long-lived bidirectional tunnel; edges cannot cache it, only carry it. What the edge contributes: TLS termination near the user, tuned backbone paths to origin, DDoS filtering in front of your socket tier, and connection fan-in. What to interrogate: idle-timeout policies (silent minute-scale timeouts are the classic mystery disconnect), concurrent-connection limits and their pricing, and what happens to established sockets during edge deploys and drains, graceful migration or mass reconnect storms.
gRPC: HTTP/2 semantics all the way
gRPC rides HTTP/2 frames with trailers, and trailers are where naive proxies die: any hop that buffers full responses or drops trailing headers breaks status propagation and streaming semantics. End-to-end HTTP/2 (or H3 with gRPC-over-QUIC where supported), no response buffering on streaming RPCs, and header-size limits that survive real metadata are the checklist. Verify with a streaming echo service through the actual edge path, not with the docs page.
A pattern gaining ground deserves mention: terminating the real-time protocol at the edge itself. Instead of tunneling every socket home, edge-compute layers increasingly hold the client connection, maintain lightweight session state, and speak efficient internal protocols to origin, fan-in at the edge, pub-sub distribution outward. Chat presence, live scores, collaborative cursors and notification fan-out fit this shape today. It converts the CDN from a pipe for your sockets into the socket layer itself, with origin connections dropping by orders of magnitude. When a vendor demos edge compute, ask about long-lived connection support in the runtime; it is the capability that separates request-rewriters from platforms.
The reconnect economics
Every long-connection estate lives one incident away from a reconnect storm: an edge restart or route shift disconnects a million clients whose retry logic then synchronizes. The mitigations belong to you as much as the vendor: jittered exponential backoff, resume tokens so reconnects skip re-auth and state rebuild, and server-side admission control. Ask providers how their maintenance drains long connections; the good ones have a story with numbers in it.
In practice
Pilot with production-shaped traffic: hold ten thousand sockets through a provider maintenance window and count survivors; stream a slow gRPC response and confirm frames arrive incrementally with trailers intact. Real-time infrastructure rewards paranoid acceptance testing precisely because its failures are user-visible instantly and dashboard-visible late.
Real-time delivery reviews include the disconnect-storm tabletop and a trailer-integrity test. Both have surprised clients.