BigData Boutique ads

Building AI on OpenSearch? Join BigData Boutique and Oracle on November 18 (12:00 PM CET) for a free, live, demo-driven webinar on OpenSearch on OCI. You'll see: - RAG and vector search running live on OCI Search with OpenSearch - How dedicated ML nodes fit in - Real-world agentic AI use cases, with lessons from building NeverBlink, our AI DBA for OpenSearch Speakers: Itamar Syn-Hershko (BigData Boutique), Piotr Kurzynoga and Irine Benoy (Oracle). Live Q&A included. Details: https://lnkd.in/eSXwH6Fd

You probably know who owns the database in your org. But do you know who can keep it safe when the defaults stop working and ignore all AI slop? "The platform team manages it" often only means they receive the alert first. Running on a managed database service doesn't solve this. AWS itself notes that even fully managed #PostgreSQL and #OpenSearch still needs workload-specific monitoring and maintenance. Same goes for Elastic Cloud, #Clickhouse Cloud and pretty much any managed database: The infrastructure is managed. The judgment isn’t. Duffel learned this through a 2-hour, 17-minute API outage. A routine partition operation collided with an anti-wraparound autovacuum. Reads and writes piled up behind a lock. Nothing about the individual operations was unusual. The danger was in how they interacted. This is the gap I see in many of our customers. The DBA role has disappeared, but the DBA work has been spread across developers, SREs, and platform engineers - especially now in the AI era. Usually without enough time or deep expertise in every database they operate. The answer isn’t necessarily hiring a full-time DBA for every engine, but every company running a production database should be able to answer: - Who catches slow degradation before customers do? - Who reviews risky schema and configuration changes? - Who owns maintenance, upgrades, capacity, and restore testing? - And who understands the strange query plan or finds a missing index at 2 a.m.? AI may help make that expertise more available, but production databases need evidence, approvals, and rollback paths. And someone on the driver seat that knows what they are doing and can avoid the AI slop that could damage the situation even further. In your company, is database expertise a defined capability, or a person everyone hopes will be online?

Like everyone, we need marketing help, so decided it'd be great to put our production GenAI experience to good use and get AI to do marketing for us. We call it #GTMHacker. It's already doing nice portions of the work for us almost autonomously. It's all about breaking big tasks to small pieces and letting highly specialized agents to bite in small pieces, and keeping great memory every step of the way. Here's how it's done: * Agents use Scrum Boards to take tasks to perform, and put new tasks as output. * Many specialized agents: for Content, SEO/GEO, Paid Ads, Social, and more. * We use strategic Goals to steer direction and serve as the northern star for any execution. * Campaigns to track mid-range strategy. * And Assets and Reports which serve as memory and output. Combining hyper-focused agents with short- and long-term memory constructs is key. --- An actual example of a time consuming task handled entirely by agents is Link Building. A research agent finds potential domains to put links on; a triage agent does a deeper research and considers, then provides feedback. Another one will draft approach email and send it. Agents will create tasks and move them in the pipe, raising a flag if human review where needed (e.g. before sending an email on his behalf). Ideally getting hundreds of new links can be done in hours (or days because we depend on a human to respond) not months. There tons more to it - from content creation (curation?) to Paid Ads and writing newsletters. It's a step by step process as we "test it in production".

We are building #GTMHacker as an experiment for building a marketing team that is fully agentic and automated. And it's working well! This is my second update. We now have a fully working Paid Ads and Content Calendar boards. For ads, agents review the current status and decide on next steps. They use good old (agent-optimized) scrum boards to communicate and delegate tasks. How do we maintain memory and situational awareness? - Agents define campaigns and launch them - Reports are generated periodically and keep memory of stats, changes and experiments - State of the art playbooks and guides are curated and used to guide the agents (we do not trust LLMs knowledge blindly) - Any review and status report looks at 1-3 reports back to understand status and avoid going in circles. - Scrum boards are part of agent outputs - tasks to do next. Some tasks will trigger the next agentic run, some runs are periodic. We are already seeing great results from this, now evaluating some improvements to the existing flows - and in parallel implementing new boards and capabilities.

ClickHouse in production: slow queries, replication lag, and schema sprawl aren't config problems — they're architecture problems. We see the same pattern repeatedly: → Queries that worked at 10M rows crawl at 1B → Replication falls behind during peak ingestion → MergeTree settings that made sense at launch now cause mutation storms Most teams try to fix this with more hardware or parameter tuning. But when the data model fights the engine, no amount of resources helps. The 3 highest-impact fixes are almost always: sort key redesign, projection strategy, and materialized view cleanup. #ClickHouse #DataEngineering #RealTimeAnalytics

Most teams building AI agents don't need AI agents. If your workflow is "parse this, validate that, post to database" - that's a pipeline. You don't need an LLM deciding what to do next. You need an if statement. Agents shine when the problem genuinely requires adaptive reasoning - when the next step depends on what you just learned and can't be predicted upfront. Here's what nobody tells you before building one: - A 5% error rate per step compounds to 40% over a 10-step chain - Over 75% of production agent systems use multiple models, not one - Only 37% of teams actually evaluate their agents in production - The no. 1 blocker isn't cost or latency - it's quality The real engineering challenge isn't making the demo work. It's the 6 months after. We wrote up everything we wish we knew before our first agent project. Link in comments 👇