https://xitter.cf/t.co/vlD2sAhx3c

localhost
Joined September 2023
Data MCP servers: don't spawn a process per tool call. Each re-pays the remote schema cost. Same chDB query, 105-column Parquet on public S3: new process per call: 3.7s 2nd call, same process: 1.0s Import + init: 0.08s. Most of the gap is schema inference, cached per process.
2
1
207
chDB retweeted
1/5 ClickHouse 26.10 can reduce hash-table memory for large GROUP BY and uniqExact queries by leaving fewer slots empty. Load factor means the fraction of allocated slots that hold keys.
2
2
6
281
chDB retweeted
Since Maple has grown into having such a big product service, and a lot of people are searching for an easy way to add instrumentation to their sideprojects I'm starting to work on making Maple Local (basically a speced down OTEL only version of Maple), easy to self host. One docker container, attach a volume and you are done. Runs on 1GB of ram, includes ingest, dashboard and query api. This is one the easiest and cheapest ways to self host your own small observability platform. Love to hear any use cases or wishes to extend the self hosted option here, will add one click deploys for various platforms soon. maple.dev/docs/self-hosting/
Maple has naturally been transforming into more than just an OTel-based observability platform, which honestly makes it harder for me to pitch and describe. It's not just a warehouse with a query engine and an MCP server. Maple deeply integrates with all your providers, works amazing for product analytics, web errors and session replays, and also lets you build alerts on literally anything. User signups down, memory usage in your pods too high, latency of a specific endpoint too high, or even alerts on slow queries or database load. The possibilities are endless. Maple also helps you improve your CI, does AI code review with all the context of your production data, and soon will include security monitoring across all your connected services, be it AWS, OpenRouter usage, Cloudflare, your email sending service etc. On top of that, full-on deep AI security scans across your codebase and your infra. With agents now, the value of having all your monitoring, analytics and observability in one place is so high. They can deeply understand how your entire stack works to improve performance, debug issues and find security issues. Still not sure how to really describe this too well and showcase the value you get from this commitment into using Maple and embracing the platform.
8
2
1
37
2,155
Agent state lives in a DataFrame, history in Parquet. One query joins both, no load step: SELECT tool, avg(ms), any(usual) FROM Python(calls) c JOIN (SELECT tool, avg(ms) usual FROM file('hist.parquet') GROUP BY tool) h USING tool GROUP BY tool chDB, 50k x 20M rows: 0.06s
2
191
Good to see @EnterpriseDB using the pg_stat_ch Postgres extension alongside OTEL and @ClickHouseDB to power observability for their Postgres customers. Validates the quality of the extensions that we launch. Btw, pg_stat_ch sits at the heart of every ClickHouse Managed Postgres instance, powering Query Insights, a fond feature for our customers. And yes, pg_stat_ch open source. github.com/ClickHouse/pg_sta… Excited to see the talk recording at @pgconfeu !
2
2
13
901
Useful pattern from ai-visibility-dataplane: 96 lines of Python serve the slice of ClickHouse's HTTP interface the app uses (queries, inserts, typed params, dedup tokens) from chDB. No Docker locally. CI runs the real clickhouse-server, same SQL. github.com/manojrevenkar12-w…
1
2
216
Agent memory that stores only the chat misses most of what the agent saw. Claude Code JSONL since July, 119k blocks in chDB. Sessions with "Traceback": user/assistant text: 22 tool results: 290 Keep tool_result rows in the same table, kind as a column. One ILIKE over both: 0.11s
2
118
A benchmark notebook that checks answers before timings: 30.4M PyPI download rows, 463,523 groups, and chDB (file queries + MergeTree), DuckDB, Polars and pandas each verified against an independent streaming Counter. All CPU stages matched. github.com/vikramaditya-tatk…
1
2
212
Subtext grounds Mexican Spanish subtitles in 103.6M subtitle pairs. When its hackathon ClickHouse Cloud trial ended, the 5 tables the site reads (~1 GB) went to Parquet inside the Cloud Run image, and chDB now serves the same SQL in-process. github.com/BrunoGupa/subtext
125
Agent rewrote your pipeline? Diff the output first: SELECT * FROM file('old.parquet') EXCEPT SELECT * FROM file('new.parquet') Run both ways. Row order is ignored, duplicates aren't. chDB, 10M rows: 3 diff rows in 3.2s. A pandas outer merge missed the duplicate.
1
66
20 people opened our dashboard at once. Every page returned a 504. A few weeks later, the same 3 CPUs serve 200 concurrent users with zero errors. @SiddarthAhhaaa walked through exactly how in our latest engineering session. The fixes, with the numbers 🧵
5
3
2
17
1,082
Prepping for the ClickHouse Certified Developer exam? This GitHub study lab runs 29 hands-on labs in a local chDB sandbox, auto-checks each answer against the sandbox's system tables or a reference query, and adds a 2-hour mock exam. github.com/bojianli2010/clic…
2
203
2023's global #podcast rankings now migrated into @ClickHouseDB from MongoDB! I'm now at about the halfway point migrating the 4.5 TB of data, and my Clickhouse + Postgres DBs are barely breaking a sweat at a combined total of 219 GB!
Now I've got 2022 global #podcast rankings migrated for Podgagement! At this pace, it might finish before the week is over and then I can ditch my 4.5 TB MongoDB database and start saving about $700/month!
1
1
5
308
m2c records every request your AI agents make, tagged by session and tool, via a mitmproxy add-on into embedded chDB. One data dir per scan, an on-disk spool so no batch is lost, 142 bytes per flow on LLM traffic, 20-35 ms queries over 1M flows. github.com/filipnyquist/m2c
92
Data agents should write pandas, not SQL. Models have seen far more DataFrame code than any SQL dialect. pandas' limit was never the API. It was holding every row in RAM. Same groupby, 10M-row 1.2 GB CSV: 7.1s, 3.9 GB in pandas. 1.2s, 1.4 GB after import chdb.datastore as pd
1
5
312
An A/B test readout with its metrics in ClickHouse SQL on embedded chDB: MergeTree tables, pandas frames loaded via Python(frame), and a CI test that fails if the SQL numbers drift from a separate pandas recount. The SQL runs unchanged on a server. github.com/nagina0205-analyt…
2
3
228
chDB retweeted
1/5 ClickHouse is getting a disk backend for its optional query cache, so saved results can survive a server restart. Useful when lots of people open the same expensive report and can tolerate a short delay before new data appears.
2
1
3
233
Good checklist via @CaioRicciuti for a local database holding state you'd miss: name the failure to survive, keep the working copy apart from the recovery copy, decide who writes, test restore on a fresh machine. chDB Durable Layer is the example. caioricciuti.com/blog/local-…
1
3
199