Research topic
Resilience
Recovery starts with an independent control path.
Design and testing guidance for containment, restoration, and control-plane recovery under pressure.
View every topicFrom the desk
38 publicationsRecovery objectives that match the cloud service
Define the business function, outage clock, recoverable data and dependency assumptions before choosing a cloud disaster-recovery architecture.
A controlled return from the SQS dead letter queue
Repair the failure, check consumer compatibility and return failed work with a bounded rate, observable stop conditions and business reconciliation.
Error budgets for controlled service degradation
Protect essential work under load while counting rejected and degraded requests against the service promise that users were actually given.
Certificate renewal under shorter validity limits
Use the public TLS issuance schedule to review authorization, renewal, deployment and independent verification of the certificate an endpoint actually serves.
The bottlenecks that shape a cloud DDoS response
Distinguish bandwidth, packet processing, connection state and application work before choosing a DDoS response or assuming the whole service path is protected.
Where DNS failover loses control of the clock
Separate authoritative routing, resolver caches, stale answers, runtime caching and existing connections when describing what DNS failover can achieve.
Keep database changes compatible with application rollback
Preserve an explicit relationship between old code and migrated state through additive changes, safe backfills, and a defined rollback window.
Stop retries from amplifying an outage
Count attempts across the complete request path, give retries a finite owner and budget, and define how repeated intent avoids duplicate side effects.
Keep encryption keys recoverable with the data they protect
Trace each encrypted recovery point to its required key, usable lifecycle state, and restore permissions before retiring cryptographic dependencies.
Make regional failover work without new infrastructure
Prepare capacity, dependencies, and the routing control path before an incident, then measure when clients reach an accepted recovery service.
Measure recovery by the service you can restore
Define application acceptance, recoverable data, and a complete timeline before treating a completed restore job as proof of recovery.
Protect backup copies from the account that runs production
Map deletion authority, retention protection, keys, and recovery identities so a surviving backup has a usable path back to service.
Qwen3.8-Flash-Next and GLM-5.3-Flash share a 3:1 long-context pattern
Both models replace most conventional attention layers with recurrent state and reserve sparse attention for periodic retrieval. Their differences lie in where they place capacity, how much neural computation they activate, and what their serving stacks must keep trustworthy.
Recover the control plane before you need it
A recovery plan that depends on the compromised identity system is an aspiration. Build and test an independent administrative path.