What Cachee Can Do for T-Mobile

T-Mobile processes billions of network events per day. Every call setup, every data session, every fraud check, every subscriber lookup touches a cache layer. Today, that layer is built on Oracle TimesTen for real-time fraud detection, CDN caching for web optimization, and streaming video optimization for bandwidth management.

Each of these systems has a latency floor defined by the technology underneath it. Oracle TimesTen — an in-memory relational database used as a writable cache — operates at 10–50 microseconds per lookup depending on query complexity and load. CDN edge caches add 5–50 milliseconds per miss. Video optimization engines add processing latency that compounds across millions of concurrent streams.

Cachee operates at 28.9 nanoseconds. That is 345x faster than TimesTen's best case and 17,300x faster than its worst case. This article examines what that performance gap means across T-Mobile's specific infrastructure, system by system.

1. Real-Time Fraud Detection: Oracle TimesTen → Cachee

What T-Mobile Uses Today

T-Mobile uses Oracle TimesTen as an in-memory writable cache for real-time fraud detection. TimesTen is a relational in-memory database that sits in front of Oracle Database, providing sub-millisecond reads for fraud scoring models. When a call is initiated or a data session is established, the fraud detection system queries TimesTen for subscriber risk scores, recent activity patterns, and known fraud indicators.

TimesTen's published latency is 5–10 microseconds for simple key-value lookups and 20–50 microseconds for more complex SQL queries with joins. In production under load — which is where fraud detection operates — latencies trend toward the higher end.

What Changes with Cachee

Cachee replaces TimesTen as the L1 cache tier for fraud detection lookups. The fraud scoring model's most frequent queries — subscriber risk score, recent call patterns, SIM swap history, known fraud device IDs — are served from Cachee's L0 hot cache at 28.9 nanoseconds instead of TimesTen's 10–50 microseconds.

Why This Matters for Fraud Catch Rate

Fraud detection is a race against time. When a fraudulent call is initiated, the carrier has a narrow window — typically under 100 milliseconds — to score the transaction, decide whether to block it, and execute the block before the call connects. Every microsecond of cache latency is a microsecond less for the scoring model to run.

At 50 µs per TimesTen lookup, a fraud check that requires 5 cache lookups (subscriber risk, device history, location check, velocity check, network pattern) consumes 250 µs — a quarter of a millisecond — just on cache reads. With Cachee, those same 5 lookups take 144.5 nanoseconds total. The fraud model gets 249,855 nanoseconds back — nearly a quarter millisecond of additional compute time for more sophisticated analysis, more features checked, higher catch rates.

For a carrier processing 1 billion fraud checks per day, the cumulative latency savings are staggering:

• TimesTen: 1B × 5 lookups × 50 µs = 250,000 seconds/day = 2.89 days of compute time per day
• Cachee: 1B × 5 lookups × 28.9 ns = 144.5 seconds/day = 2.4 minutes
• Saved: 2.89 days of compute time per day, every day

Oracle License Elimination

Oracle TimesTen is licensed per CPU core at approximately $47,500 per core per year (list price). A production deployment spanning 20 cores costs $950,000 per year in licensing alone — before hardware, support, and DBA costs. Cachee is a single Rust binary that runs as a sidecar container. No per-core licensing. No Oracle DBA required. No RAC complexity.

2. CDN Edge Cache: Reducing Origin Fetches

What T-Mobile Uses Today

T-Mobile uses CDN caching for web infrastructure, optimizing static assets like JavaScript, stylesheets, images, and API responses to reduce load times for T-Mobile.com, the T-Mobile app, and customer self-service portals.

CDN edge caches have inherent limitations: cache misses require origin fetches (50–200ms round trip), cache invalidation propagates slowly across edge nodes (seconds to minutes), and personalized content — account balances, usage data, plan details — is often uncacheable because CDN caches are shared across users.

What Changes with Cachee

Cachee deploys at the origin as an L1 cache in front of T-Mobile's backend services. When a CDN edge cache misses, the origin fetch hits Cachee at 28.9 nanoseconds instead of a database at 5–50 milliseconds. This is particularly valuable for personalized content that CDN edges can't cache:

• Account balance lookups: 28.9 ns from Cachee vs 15 ms from the account database
• Plan details and entitlements: 28.9 ns vs 8 ms from the provisioning system
• Usage data (minutes, data, messages): 28.9 ns vs 20 ms from the billing system
• Promotional offers personalized to subscriber: 28.9 ns vs 12 ms from the marketing engine

With 100 million app sessions per day and an average of 10 personalized data fetches per session, that's 1 billion origin requests per day. At 15 ms average database latency, that's 4.17 hours of cumulative latency per day. Cachee reduces that to 28.9 seconds.

3. Subscriber Data Lookup: The Core Network Cache

The Scale

T-Mobile has over 120 million subscribers. Every network event — call setup, data session, handoff, roaming authentication — requires a subscriber profile lookup. The Home Subscriber Server (HSS) and Unified Data Management (UDM) systems handle these lookups, typically backed by an in-memory database or distributed cache.

At 5G speeds, the latency budget for subscriber lookups is measured in single-digit milliseconds. A subscriber profile lookup that takes 5 ms on a traditional database is acceptable on 4G but becomes a bottleneck on 5G where the air interface latency is under 1 ms. The cache layer in front of the subscriber database is the critical path.

What Changes with Cachee

Cachee serves subscriber profiles from L0 at 28.9 nanoseconds. For the 10 million most active subscribers (top 8% by session count), their profiles are in L0 at all times. For the remaining 110 million, profiles are in L1 (DashMap at 59 ns on Graviton4) and promote to L0 on first access.

The impact on 5G session setup time is significant. A 5G NR call setup that currently takes 15 ms with a 5 ms subscriber lookup contribution drops to 10.000029 ms with Cachee — effectively eliminating the cache contribution entirely.

4. Video Streaming Optimization

What T-Mobile Uses Today

T-Mobile employs streaming video optimization to manage network load, which includes caching popular video segments to decrease bandwidth usage. During peak hours, the same popular content — trending videos, live sports highlights, viral clips — is requested by thousands of subscribers simultaneously.

What Changes with Cachee

Video manifest files (HLS/DASH) and popular segment metadata are cached in Cachee at 28.9 ns. When a subscriber requests a video, the manifest lookup — which determines which quality level, CDN edge, and segment to serve — comes from Cachee instead of the video optimization engine.

For a carrier serving 50 million video sessions per day with an average of 30 manifest lookups per session:

• Without Cachee: 1.5B lookups × 2 ms = 3,000,000 seconds/day = 34.7 days of latency
• With Cachee: 1.5B lookups × 28.9 ns = 43.35 seconds/day
• Saved: 34.7 days of latency per day — more than the day itself

5. Google Cloud Analytics Acceleration

What T-Mobile Uses Today

T-Mobile has migrated data analytics and infrastructure to Google Cloud, using BigQuery and other GCP services to predict usage trends and accelerate analytics. These analytics queries scan large datasets and return aggregated results that are used by business intelligence dashboards, network planning tools, and customer experience systems.

What Changes with Cachee

Analytics results are highly cacheable. The top 100 dashboard queries at T-Mobile are probably the same queries run by the same people every morning. Without caching, each query hits BigQuery ($5 per TB scanned). With Cachee, repeated queries serve cached results at 28.9 ns.

At $5/TB and an average of 500 GB scanned per common query, each query costs $2.50. If 200 users run the same 20 queries each morning, that's 4,000 BigQuery scans costing $10,000/day. With Cachee serving 90% from cache: $1,000/day. $3.28 million saved per year on BigQuery alone.

6. RevTech Big Data Integration

T-Mobile partners with RevTech for big data processing. These batch and streaming data pipelines produce derived datasets — subscriber segments, churn risk scores, propensity models — that are consumed by downstream applications. The output of these pipelines is a natural caching target.

Cachee sits between the RevTech pipeline output and the consuming applications. When a churn risk score is computed for a subscriber, it's written to Cachee once and served at 28.9 ns to every application that needs it — marketing automation, retention outreach, customer care screens, network prioritization. Without Cachee, each application queries the data lake independently at 50–200 ms per lookup.

The Total Financial Impact

The Deployment Path

The fastest path into T-Mobile's infrastructure is through fraud detection. Oracle TimesTen is the incumbent. The performance gap is 345x. The license cost is $950K per year. The replacement is a single Rust binary that deploys as a sidecar container — no Oracle DBA required, no RAC complexity, no per-core licensing.

Phase 1: Deploy Cachee alongside TimesTen as an L0 tier. Hot fraud detection keys serve from Cachee at 28.9 ns. TimesTen becomes L2 for cold lookups. Zero risk — TimesTen stays in place as fallback. Measure the hit rate and latency improvement over 30 days.

Phase 2: Once L0 hit rate exceeds 95%, begin routing 100% of fraud lookups through Cachee. TimesTen handles only cold-start scenarios and historical queries.

Phase 3: Decommission TimesTen for fraud detection. Redeploy Cachee across CDN origin, subscriber lookup, and video optimization. Each deployment follows the same L0-first pattern.

The conversation starts with one question: "What does your fraud detection stack look like under peak load? Where does latency start hurting decision quality?"

Try the live speed test · Run your numbers · 5G Telecom vertical · Start free trial