Case study · Quick commerce · Real-time data
Helped India's leading quick-commerce platform reduce average delivery time by 30%.
Quick commerce · 10-minute delivery
Batch pipelines meant every metric arrived a day late. We rebuilt the data path around change data capture and got it to fifteen minutes.
15 min
Analytics freshness, down from 1 day
5x
Improvement in overall operational process
30%
Improved driver availability metrics
40%
Lower ETL tool cost
The opportunity
Ten-minute delivery promises, running on day-old data.
This platform delivers 2,500+ products in ten minutes from delivery centres across ten locations. Its data pipelines were built in batch mode, so significant lag sat between what happened and what anyone could see-which is a hard way to run an operation measured in minutes.
That delay pushed into everything downstream: estimated arrival times and demand-supply balance could not be computed in near real-time, packer movement inside the warehouse was hard to track, and inventory visibility lagged far enough to compromise whether an order could be fulfilled at all.
Batch data pipelines
Initial pipelines ran in batch mode, producing significant data lag.
ETA and demand-supply lagged
The right driver metrics could not be computed in near real-time.
Packer movement untracked
Monitoring packer activity inside the warehouse was difficult.
Inventory visibility behind
Near-real-time stock levels were unavailable, compromising fulfilment.
The impact
A day of latency became fifteen minutes.
Change data capture, not batch
Debezium connectors on MSK Connect capture changes from Aurora PostgreSQL as they happen.
Five-minute minimum frequency
Incremental exports from S3 into Redshift run per table, as often as every five minutes.
Driver availability up 30%
Scheduled aggregation queries surface stock levels and resource allocation while they still matter.
Data mesh across 500TB
Workloads segregated by team so operations, revenue and analytics each build what they need.
The stack
What it runs on.
- Change capture
- Debezium connectors on Amazon MSK Connect
- Source and landing
- Aurora PostgreSQL, Amazon S3 destination connectors
- Warehouse
- Amazon Redshift with a data mesh architecture, ~500TB
- Orchestration
- Managed Apache Airflow, custom in-house ETL framework
The road ahead
The data is current. Next is what acts on it.
With freshness at fifteen minutes and workloads separated by team, the next gains come from what each team builds on top-tighter ETA models, and allocation that responds to the signal rather than reporting it.
- 01
Tighter ETA modelling
- 02
Automated resource allocation
- 03
Per-team analytics expansion
Talk to us
Bring one process. Leave with an outcome architecture.
30 minutes on one function, mapped to the architecture behind it. No slides.