TripAccordOpen the planner

Engineering notes · November 2026

What does a shared trip actually cost?

The organizer reads the chat once. A dozen travelers may open the itinerary. What do those two jobs cost?

The cost of keeping a plan online

At 2,000 travelers and twelve visits each, a 750 KB visit adds up to 18 GB a month. These monthly estimates cover the app and its pages, in US dollars before inference.

Configuration 50 travelers 2,000 travelers
Akamai VM, SQLite and Object Storage $10.00 $10.00
Google Cloud Run and Cloud SQL $11.61 $13.99
Azure Container Apps and MySQL $14.75 $15.33
AWS Lambda and RDS PostgreSQL $20.03 $20.85
Akamai Kubernetes, one worker, load balancer and SQL $43.00 $43.00

On Akamai’s $10 setup, the operator handles backups and recovery. Google and Azure include a managed database and assume the app’s monthly compute allowance is available. A second worker takes Akamai Kubernetes to $55.

Where Akamai changes the bill

Share enough photos to send 1,000 GB in North America and the gap widens. That traffic fits Akamai’s included transfer. Estimated network charges reach $111.64 on Google’s Premium Tier, $78.30 on Azure and $81 on AWS.

Reading the chat

A Seattle CPU worker took 34 minutes to fetch and verify the 4-bit Qwen model. Restarting with cached weights took 9.6 seconds. Both trip conversations reached the ten-minute limit without a complete plan. The trial cost $0.18.

What happens to a cover photo?

The app sends a short-lived link. Seattle Object Storage sends the picture. All 50 downloads from British Columbia matched the original, with a median of 50 ms to the first byte. A backup in the same bucket restored the itinerary and its cover.

What survives a restart?

Switching containers on a single $5 Seattle VM served all 422 requests made during the switch. On Kubernetes, replacing both readers in Seattle, Frankfurt and Tokyo served all 998 requests during the three updates. Each response matched the approved version.

On Akamai’s $16/month managed PostgreSQL node, a dump and restore of the itinerary and cover took 1.96 seconds. All 100 checks on the restored data matched.

How far away is that link?

Each location opened the Seattle itinerary and cover 50 times apiece over fresh HTTPS connections. Choose the median for a typical wait or the 95th percentile for a slower one.

Compare response times
Median time to first byte, milliseconds
Reader Itinerary Cover
Seattle 24 ms 98 ms
British Columbia 56 ms 172 ms
Tokyo 289 ms 615 ms
Frankfurt 440 ms 816 ms
95th percentile time to first byte, milliseconds
Reader Itinerary Cover
Seattle 49 ms 140 ms
British Columbia 79 ms 238 ms
Tokyo 373 ms 671 ms
Frankfurt 467 ms 851 ms

Read the study, assumptions and measured results →