About the service
I design and operate distributed wide-column databases on Apache Cassandra and ScyllaDB, and I always start with one thing: your queries. Before spinning up a node, I map all read and write paths your app actually needs, then build a data model around those access patterns. In column-oriented storage, the table follows the query, so I deliberately denormalize, choose partition keys with even ring distribution, and pick clustering columns that land hot reads in one partition.
Then I tune the cluster to hold real traffic. I set replication factor and consistency levels to your budget for reliability and latency, cap partition size, choose compaction strategy for your load — time series, event log, or high-churn key-value. I keep write pipelines at tens of thousands of ops/sec without stalls during flushes and compaction, and tail latency predictable: wide partitions, tombstones, and hot shards I catch early. On ScyllaDB I use shard-per-core architecture and extract significantly more throughput from the same hardware.
I also own the tasks that break clusters at 3 AM: repair scheduling, node replacement, cross-DC replication, capacity planning, and migrations from relational databases that stopped scaling. On the way out — a database that absorbs spikes and grows by adding nodes.
— Data modeling from queries and partition design
— Replication, consistency, and compaction tuning
— High write throughput and predictable latency
— Repair, node replacement, and multi-DC setup
Then I tune the cluster to hold real traffic. I set replication factor and consistency levels to your budget for reliability and latency, cap partition size, choose compaction strategy for your load — time series, event log, or high-churn key-value. I keep write pipelines at tens of thousands of ops/sec without stalls during flushes and compaction, and tail latency predictable: wide partitions, tombstones, and hot shards I catch early. On ScyllaDB I use shard-per-core architecture and extract significantly more throughput from the same hardware.
I also own the tasks that break clusters at 3 AM: repair scheduling, node replacement, cross-DC replication, capacity planning, and migrations from relational databases that stopped scaling. On the way out — a database that absorbs spikes and grows by adding nodes.
— Data modeling from queries and partition design
— Replication, consistency, and compaction tuning
— High write throughput and predictable latency
— Repair, node replacement, and multi-DC setup
