Advanced everything in Databases & Architecture
NVIDIA's AVO harness ran Claude Opus 5 autonomously for seven days straight
NVIDIA paired Claude Opus 5 with a new harness called AVO to run long-horizon autonomous tasks, including a seven-day GPU kernel optimisation run and a separate reasoning benchmark. AVO uses persistent memory and a supervisor process so the agent keeps working past a single context window instead of restarting.
AI code volume is pushing code review earlier, ahead of the pull request
AI now produces more code than humans can realistically review, with Meta's lines of code per human-landed diff reportedly up 106% in a year. One response argues review should move earlier, into pairing and design sessions, leaving the pull request to catch formatting and known security issues.
Tuning HNSW for recall vs latency
ANN trades exactness for speed, and HNSW knobs M and efConstruction shape graph quality while efSearch trades query latency for recall at runtime.
TSM-Tree vs LSM-Tree storage engines
Both buffer writes in memory and flush sorted immutable files, but TSM organizes by series and time with columnar, heavily compressed blocks tuned for ordered appends and range scans.
Iceberg vs Delta Lake metadata and ACID
Iceberg uses a tree of metadata and manifest files with atomic pointer swaps and optimistic concurrency; Delta uses an ordered transaction log of JSON commits with optimistic concurrency.
The small files problem in data lakes
Many tiny files create per-file overhead and metadata pressure, hurting scans; fix via compaction, batching writes, and tuning partitioning.
Multi-region active-passive DR with Aurora
Async global replication to a passive region, promote and repoint traffic on failover, and fence the old primary to prevent split-brain.
Aurora vs Spanner architecture contrast
Aurora is single-writer with a shared distributed log-based storage and quorum, scaling reads; Spanner shards data with Paxos and TrueTime for global writes.
Fixing an ORM's inefficient aggregation query
Drop to raw SQL or a view for the heavy report, or restructure the ORM query and add indexes. Raw SQL is fast but couples to the schema; tuning keeps portability.
Unit of Work / Session pattern in ORMs
The Unit of Work tracks new, dirty, and deleted objects, then flushes them as one batched transaction at commit.
Defense-in-depth against SQL injection
Beyond parameterization, apply least-privilege accounts, stored procedures, input allowlisting, and monitoring to shrink blast radius.
Diagnosing degradation with normal CPU and memory
When CPU and memory look fine, sessions are waiting, not computing; examine wait statistics, lock and latch contention, I/O waits, and buffer pool hit ratio.
Mitigating a database shard hot spot
Short-term, add read replicas or cache the hot keys; long-term, fix the partition key with hashing, salting, or finer-grained splitting.
Split-brain, consensus, and quorum
Split-brain is two nodes both believing they are leader during a partition; Raft/Paxos require a majority quorum to elect a leader and commit, so the minority side cannot make progress.
Durable write path in a sharded KV store
Route by key hash to the shard leader, append to WAL and fsync, replicate to two followers, ack on quorum, then confirm.
What is an OLAP cube and its operations?
A cube pre-aggregates measures across dimensions; operations are slice, dice, drill-down, roll-up, and pivot.
Why separate storage and compute in a cloud warehouse?
Data lives in cheap shared object storage while independent compute clusters scale separately, enabling elastic, concurrent, isolated workloads and pay-per-use.
What consistency do you sacrifice in an AP system?
You give up linearizability and often sequential consistency, accepting stale reads and conflicts, then mitigate with quorums, vector clocks or CRDTs, and…
How do you keep consistency without multi-document transactions?
A Saga runs a sequence of local transactions, each with a compensating action to undo on failure, coordinated via choreography or orchestration.
How does a hash join handle memory overflow?
The build table is partitioned by hash and spilled to disk, then probe rows are partitioned the same way, and pairs are joined per partition.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles