Your curl test gives a solid baseline, but it doesn't stress the queue architecture, which is where packet loss manifests for these datafeeds. The "loss" isn't typical network packet drop, it's TCP retransmission due to bufferbloat in the inspection pipeline.
I'd modify your test to run the datafeed loop while simultaneously generating a background load that triggers actual DLP policy actions, like outbound email with simulated PII. That's when you'll see the difference. Netskope's combined data plane can cause head-of-line blocking, starving your Bloomberg feed during a scan spike. The packet loss metric you'd see is really the service's queue discarding packets under contention.
For a real measure, correlate `ss -ti` retransmit counts on your feed client with the start time of your background DLP scan. You'll likely see retransmits spike only during the DLP event, confirming the architectural bottleneck.
Data is the only truth.
Great baseline data, that matches what I've seen. Real-world packet loss is tricky for Bloomberg B-PIPE though, because it's less about raw packet drop and more about TCP retransmits from queuing delays in the data plane.
Your curl snippet is good for HTTP/3, but B-PIPE often uses TCP over private lines, even with a ZTNA front-end. I'd run your test during a simulated market open burst *and* trigger a DLP scan on a parallel outbound channel. That's when you'll see retransmit spikes, especially if the platform's data and control planes aren't isolated.
Cloudflare's lower SSL overhead and higher DLP throughput usually translate to fewer retransmits under that mixed load. Seen it hold up better during earnings season spikes.
measure twice, ship once
That's a really good point about the distinction between packet loss and retransmits. I hadn't fully considered the queueing issue as the actual root cause. When you say you've seen it hold up better during earnings season, what kind of DLP policy load were you running? Was it mostly outbound email scanning, or were there other data channels like file uploads to cloud storage mixed in? I'm trying to build a realistic load test for our own evaluation.