Help: Our trace volume exploded and now our bill is insane. What to do?

Alright, who else got blindsided by last month's Freeplay bill? We've been running their tracing for about six months, ingesting from our LangChain apps and a couple of custom Pinecone rag pipelines. Volume was steady, maybe 50k spans/day. Then we shipped a new feature that does recursive query decomposition. Suddenly we're at 3 million spans a day and our bill jumped from a couple hundred bucks to over five grand. My CFO is having a stroke.

I've spent the last two days neck-deep in the docs and our own code, and I've found a few levers, but I want to see what the rest of you crusty pipeline folks have done. The pricing per span is brutal at scale, and while the UI is nice, I'm not paying for a pretty graph.

Here's what I'm looking at so far:

* **Sampling at the source:** This is the big one. We're using the OpenTelemetry SDK. Pushing everything is for startups with VC money. I'm testing a head-based sampler.
```python
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.sampling import TraceIdRatioBasedSampler

# Sample down to 25% for non-critical user-facing paths
tracer_provider = TracerProvider(sampler=TraceIdRatioBasedSampler(0.25))
```
The trick is applying this discriminatively. Health checks, internal cron jobs? Sample at 1%. Key user workflows? Maybe 100%.

* **Span attribute diet:** We were lazy and dumping entire object JSON into attributes like `llm.prompts`. Freeplay charges by attribute count per span. I wrote a processor to strip out everything but the template name and token counts.
```python
# Pseudo-code for an OTel span processor
def attribute_trimmer(span):
if "llm.prompts" in span.attributes:
# Replace fat JSON with just a hash or identifier
span.attributes["llm.prompt_template"] = extract_template_name(span.attributes["llm.prompts"])
del span.attributes["llm.prompts"]
```

* **Aggregate metrics export:** For pure cost monitoring (token usage, latency p99), we're bypassing traces altogether and pushing directly to our own Prometheus. Cheaper to alert off that.

My questions for the group:
1. **Downsampling after the fact:** Can you configure Freeplay to only *store* a sample, even if you send everything? I haven't found this, and I suspect it's by design.
2. **Batch exports for analysis:** If we sample heavily, we lose outliers. Is anyone doing a dual export? Send 1% to Freeplay for the UI, and send 100% to a cheap object store (S3, GCS) for later bulk analysis if something goes wrong?
3. **Negotiated pricing:** Has anyone with serious volume (>5M spans/day) had luck talking to sales about a committed-use discount, or are we better off architecting our way out of the problem?

I like the product, but I'm not running a charity for their cloud bill. The old adage holds: "It's not a data pipeline if you can't control the faucet." What are your control valves?

-- old salt