Accordion desk
Answers skew toward logistics, telemetry, and what finance sees on invoices.
No. Week one runs against our shared lab; you only need Docker-capable hardware and VPN access we provision after enrollment confirmation.
See the Refund pathway page for eligibility windows, partial refund cases, and processing timelines under Korean consumer guidance.
Seats are per named participant so feedback stays attributable; we can invoice multiple cost centers on request.
Only anonymized aggregate timings for lab UX improvements; raw logs never leave the Hwaseong perimeter without a signed addendum.
Primary materials are English with Korean glossaries; facilitators answer conceptual questions in either language during office hours.
Quarterly unless a vendor security bulletin forces sooner; we publish change logs so repeat learners know what drifted.
Case ledgers
Industrial HVAC OEM
Challenge: 5-minute sensor bursts overwhelmed a legacy JDBC sink during monsoon season spikes.
Approach: Cohort rebuilt ingestion with Spark structured streaming, adaptive batching, and Iceberg compaction windows aligned to maintenance crews.
Outcome: Median ingest latency dropped from 3.2 minutes to 18 seconds without enlarging the SQL warehouse footprint.
Regional payments network
Challenge: Skewed joins across two countries produced executor hotspots whenever campaign traffic doubled.
Approach: Teams applied salting patterns from the Mesh Foundry labs plus AQE tuning playbooks shared in class.
Outcome: p95 stage duration stabilized across 14 consecutive nightly runs; finance reviewers received reproducible DAG snapshots.
Consumer finance
Challenge: Regulators requested immutable feature snapshots while marketing still wanted weekly refreshes.
Approach: Delta time travel + branch snapshots modeled in the Lake Lathe module, with legal sign-off templates included.
Outcome: Audit trail satisfied external counsel while data science retained weekly sandbox branches.
Cold chain logistics
Challenge: Batch scoring missed port strikes because streaming signals were siloed from the nightly Spark job.
Approach: Merged Kafka topics into a guarded bronze layer using the Stream Sentry exercises as scaffolding.
Outcome: ETA model refresh cadence moved from 36 hours to 6 hours without breaching SLA on driver notifications.