<p><strong>TLDR:</strong> pg_clickhouse 0.2.0 ships HTTP result set streaming that drops peak memory from 600MB+ to under 86MB on large queries. ClickHouse is also pushing into Japan with a Langfuse AI observability launch and a new partner program. On the cautionary side, an AI agent caused a production data loss incident that’s making the rounds as a reminder to harden your backup strategy. Neo4j content is circulating around graph database fundamentals and agentic AI use cases.</p>
<hr>
<h2>pg_clickhouse 0.2.0 ships streaming and pushdown improvements</h2>
<p>The PostgreSQL extension for querying ClickHouse got a meaningful update. HTTP result set streaming now buffers around 50MB at a time instead of loading full result sets into memory, cutting peak usage from 600MB+ to under 86MB on datasets like NYC taxi data. The release also adds pushdown for JSONB operators (<code>-&gt;</code>, <code>-&gt;&gt;</code>, <code>jsonb_extract_path</code>), SQL date/time value functions with proper timezone translation, and array functions like <code>array_cat</code> and <code>string_to_array</code> mapped to their ClickHouse equivalents. If you’re running federated queries across Postgres and ClickHouse, the memory improvement alone is worth the upgrade.</p>
<hr>
<h2>AI agent causes production data loss</h2>
<p>A cautionary post is circulating about an AI agent wiping production data. Details are thin, but the core message is straightforward: if you’re giving agents write access to anything that matters, your backup and recovery posture needs to be airtight before that happens, not after. Worth a read if you’re anywhere near agentic workflows touching real databases.</p>
<hr>
<h2>ClickHouse launches Langfuse AI observability in Japan</h2>
<p>ClickHouse Japan is now offering Langfuse, an AI observability platform, out of the AWS Tokyo region. It records execution steps for AI agents and LLM applications, handles real-time quality scoring, and tracks token usage against ClickHouse’s columnar backend. The stack also includes LibreChat for the frontend layer. Separately, ClickHouse launched its global partner program “House Mates” in Japan, with three tiers and co-marketing support aimed at enterprises building on real-time data infrastructure.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Neo4j and Cypher intuitiveness:</strong> Jennifer Reif demos how developers naturally reinvent Cypher’s arrow notation when describing graph relationships, which says something about how well the query language maps to how people actually think about connected data.</li>
<li><strong>Graph databases for exploratory analysis:</strong> A Neo4j segment uses the Panama Papers as a concrete example — the argument is that graph databases let relationships emerge after ingestion, while SQL requires you to define them upfront, which matters a lot when you don’t know what you’re looking for yet.</li>
<li><strong>Neo4j and agentic AI:</strong> Reif introduces the idea of a “context graph” for agentic systems — using a graph to store business decisions, process reasoning, and chat memory so agents can share context across sessions rather than starting cold each time.</li>
</ul>


Query optimization, schema design, and data store coverage across PostgreSQL, MongoDB, Redis, and more.

Databases Digest

pg_clickhouse 0.2.0 introduces HTTP result set streaming that reduces peak memory usage from 600MB+ to under 86MB on large queries, along with JSONB operator pushdown and array function mappings. A cautionary incident highlights an AI agent wiping production data, underscoring the need for robust backup strategies before granting agents write access. ClickHouse is expanding in Japan with a Langfuse AI observability offering on AWS Tokyo and a new partner program. Neo4j content covers graph database fundamentals, Cypher query language intuitiveness, and using context graphs for agentic AI memory across sessions.

pg_clickhouse cuts memory 86% with streaming, AI agent wipes prod data