Insights — Crimson UmbrellaData engineering, software and applied AI, written by the engineers who do the work. Pipelines, retrieval, evaluation, cost and the decisions in between.https://crimsonumbrella.com/en-gbA read-only Postgres role is the whole MCP security modelhttps://crimsonumbrella.com/insights/a-read-only-role-is-the-mcp-security-model/https://crimsonumbrella.com/insights/a-read-only-role-is-the-mcp-security-model/The GRANT is the only control a tool-calling agent can't argue with: what it stops, the 11.7-minute scan it doesn't, and the failure it never reaches.Thu, 29 Oct 2026 00:00:00 GMTmcpagentspostgrescontact@crimsonumbrella.com (Yunik Maharjan)A soft hyphen becomes a hyphen in three code pages out of fifteenhttps://crimsonumbrella.com/insights/a-soft-hyphen-and-a-patched-server/https://crimsonumbrella.com/insights/a-soft-hyphen-and-a-patched-server/CVE-2024-4577 turns on one record in a Microsoft mapping table, and the versions everyone calls patched carry a fix PHP deleted eight days after shipping it.Thu, 29 Oct 2026 00:00:00 GMTsecurityvulnerabilitiescontact@crimsonumbrella.com (Yunik Maharjan)We put the index break-even at 0.67% of the table, and the second feature is where you find outhttps://crimsonumbrella.com/insights/ai-agents-and-the-second-feature/https://crimsonumbrella.com/insights/ai-agents-and-the-second-feature/An AI-built prototype arrives with one index per table, a secret in the bundle and no schema history. Where the break-even sits, and what to repair.Thu, 29 Oct 2026 00:00:00 GMTdatabasespostgresarchitecturecontact@crimsonumbrella.com (Yunik Maharjan)The orchestrator line is five recoveries a month, not fifty pipelineshttps://crimsonumbrella.com/insights/airflow-dagster-or-the-cron-you-already-have/https://crimsonumbrella.com/insights/airflow-dagster-or-the-cron-you-already-have/What Airflow or Dagster is worth at 5, 50 and 500 pipelines, costed in operator minutes, and the threshold below which we won't sell a build.Thu, 29 Oct 2026 00:00:00 GMTorchestrationpipelinescostcontact@crimsonumbrella.com (Saurav Raj Joshi)Thirty test cases cannot tell ninety per cent from eightyhttps://crimsonumbrella.com/insights/build-the-evaluation-set-first/https://crimsonumbrella.com/insights/build-the-evaluation-set-first/A working demo tells you nothing about your next change. How to build an evaluation set from real questions, score retrieval separately from the answer, calibrate the judge, and gate every prompt change in CI.Thu, 29 Oct 2026 00:00:00 GMTevaluationllmcontact@crimsonumbrella.com (Yunik Maharjan)A bigger warehouse breaks even at 3.6x, or at 2.1x if you change one defaulthttps://crimsonumbrella.com/insights/capacity-is-arithmetic-not-instinct/https://crimsonumbrella.com/insights/capacity-is-arithmetic-not-instinct/A calculator for warehouse credits run on numbers you can state: credits per query-run, the idle tail, and the 600-second default that sets the break-even.Thu, 29 Oct 2026 00:00:00 GMTcostwarehousesnowflakecontact@crimsonumbrella.com (Saurav Raj Joshi)irm is an alias, not a program: four Windows install errors and the shell each one nameshttps://crimsonumbrella.com/insights/claude-code-on-windows-the-shell-is-the-problem/https://crimsonumbrella.com/insights/claude-code-on-windows-the-shell-is-the-problem/Why the Claude Code install line fails in Command Prompt, the four errors that name your shell, and the 599-build gap before WSL 2 exists.Thu, 29 Oct 2026 00:00:00 GMTagentsllmcontact@crimsonumbrella.com (Yunik Maharjan)Write payments code as if every request happens twicehttps://crimsonumbrella.com/insights/code-that-moves-money/https://crimsonumbrella.com/insights/code-that-moves-money/Five things payment code needs from the first commit: an idempotency key scoped to an intent, integer minor units, an append-only ledger, a replay-safe webhook handler and a nightly reconciliation against the provider.Thu, 29 Oct 2026 00:00:00 GMTpaymentsreliabilitycontact@crimsonumbrella.com (Yunik Maharjan)A tenfold lower EPSS threshold buys 1.54x the coverage for 8.5x the workhttps://crimsonumbrella.com/insights/epss-over-cvss-when-the-backlog-is-four-hundred/https://crimsonumbrella.com/insights/epss-over-cvss-when-the-backlog-is-four-hundred/EPSS scores how likely exploitation is in 30 days; CVSS scores how bad it would be. The arithmetic for a 400-item backlog, and the threshold we defend.Thu, 29 Oct 2026 00:00:00 GMTsecurityvulnerabilitiescontact@crimsonumbrella.com (Yunik Maharjan)enable.idempotence covers one producer session, not exactly-oncehttps://crimsonumbrella.com/insights/exactly-once-is-a-transaction-boundary/https://crimsonumbrella.com/insights/exactly-once-is-a-transaction-boundary/What enable.idempotence guarantees in Kafka, what a transaction boundary adds, and how to derive transaction.timeout.ms rather than guess at it.Thu, 29 Oct 2026 00:00:00 GMTstreamingkafkareliabilitycontact@crimsonumbrella.com (Yunik Maharjan)Your eight-hour refresh is moving data that didn't changehttps://crimsonumbrella.com/insights/full-reload-moves-data-that-did-not-change/https://crimsonumbrella.com/insights/full-reload-moves-data-that-did-not-change/Why full reloads get slow, how log-based change data capture fixes it, what has to be in place in the source database before it is safe, and when reloading the lot is still the right answer.Thu, 29 Oct 2026 00:00:00 GMTcdcpipelinescontact@crimsonumbrella.com (Saurav Raj Joshi)A 24-hour idempotency key covers 3.3% of the replay windowhttps://crimsonumbrella.com/insights/idempotency-is-a-storage-decision/https://crimsonumbrella.com/insights/idempotency-is-a-storage-decision/Idempotency key retention derived from the counterparty's published replay horizons: 24 hours outbound, 35 days inbound, and the storage arithmetic for both.Thu, 29 Oct 2026 00:00:00 GMTpaymentsreliabilitypostgrescontact@crimsonumbrella.com (Yunik Maharjan)Container queries dropped six of seven containment effects in 2024https://crimsonumbrella.com/insights/media-queries-answered-the-wrong-question/https://crimsonumbrella.com/insights/media-queries-answered-the-wrong-question/Dimensional query containers stopped applying layout containment in 2024: what container-type still does, where the specifications disagree, and when clamp() wins.Thu, 29 Oct 2026 00:00:00 GMTfrontendperformancearchitecturecontact@crimsonumbrella.com (Sahil Shrestha)A year of 4,800 Microsoft salaries is four and a half days of its property and equipment spendinghttps://crimsonumbrella.com/insights/microsoft-cut-four-thousand-eight-hundred/https://crimsonumbrella.com/insights/microsoft-cut-four-thousand-eight-hundred/What a 4,800-role cut and $39,367m of property and equipment spending do and do not tell a company buying engineering, with the arithmetic shown.Thu, 29 Oct 2026 00:00:00 GMTeconomicscareerscostcontact@crimsonumbrella.com (Saurav Raj Joshi)Most monolith splits fail at the databasehttps://crimsonumbrella.com/insights/monolith-splits-fail-at-the-database/https://crimsonumbrella.com/insights/monolith-splits-fail-at-the-database/What a service split actually fixes, why a shared schema leaves you with a distributed monolith, and which service to extract first.Thu, 29 Oct 2026 00:00:00 GMTarchitecturedatabasescontact@crimsonumbrella.com (Yunik Maharjan)Most of what gets sold as an agent should be a workflowhttps://crimsonumbrella.com/insights/most-agents-should-be-workflows/https://crimsonumbrella.com/insights/most-agents-should-be-workflows/The difference between a workflow and an agent, costed end to end on an invented system called Lamplighter: 2.4x on a good run, 35x on a bad one, and the error rate that actually decides the build.Thu, 29 Oct 2026 00:00:00 GMTagentsarchitecturecontact@crimsonumbrella.com (Yunik Maharjan)Native or cross-platform, decided by hardware then hiringhttps://crimsonumbrella.com/insights/native-or-cross-platform/https://crimsonumbrella.com/insights/native-or-cross-platform/The parts of a mobile app that still have to be native, what a native module costs across a year, and why hiring only gets a vote once hardware has abstained.Thu, 29 Oct 2026 00:00:00 GMTmobilecontact@crimsonumbrella.com (Rujen Shakya)Python did not win on syntax: four runtime boundaries and a 10,000x gap between two defaultshttps://crimsonumbrella.com/insights/python-did-not-win-on-syntax/https://crimsonumbrella.com/insights/python-did-not-win-on-syntax/The four runtime boundaries a Python data stack crosses, what each publishes as a default, and where the interpreter lock actually costs you.Thu, 29 Oct 2026 00:00:00 GMTpythonarchitectureperformancecontact@crimsonumbrella.com (Saurav Raj Joshi)Your replication slot has 39.8 hours of runway, and the copy was never the riskhttps://crimsonumbrella.com/insights/replication-slot-is-what-breaks/https://crimsonumbrella.com/insights/replication-slot-is-what-breaks/The replication slot is what breaks when you stream Postgres to a lakehouse: a stopped consumer pins write-ahead log on the source until the volume fills.Thu, 29 Oct 2026 00:00:00 GMTcdcpostgrespipelinescontact@crimsonumbrella.com (Saurav Raj Joshi)Most fine-tuning requests are really retrieval problemshttps://crimsonumbrella.com/insights/retrieval-before-fine-tuning/https://crimsonumbrella.com/insights/retrieval-before-fine-tuning/Retrieval changes what a model can see; fine-tuning changes how it answers. A diagnostic for telling which problem you have, and the arithmetic for deciding when a lexical index is worth adding.Thu, 29 Oct 2026 00:00:00 GMTragllmcontact@crimsonumbrella.com (Yunik Maharjan)The crossover is 77,000 tokens on Sonnet 5, and the fine-tune that cost $2.84 pays back in 18.7 yearshttps://crimsonumbrella.com/insights/retrieval-long-context-or-fine-tuning/https://crimsonumbrella.com/insights/retrieval-long-context-or-fine-tuning/Corpus size decides between retrieval and a long context window. Fine-tuning sits on another axis. The crossover is derived twice; the fine-tune never pays.Thu, 29 Oct 2026 00:00:00 GMTragllmcostcontact@crimsonumbrella.com (Yunik Maharjan)Row-level security is free until a 0.1% tenant reads 40 million rowshttps://crimsonumbrella.com/insights/row-level-security-and-the-planner/https://crimsonumbrella.com/insights/row-level-security-and-the-planner/A Postgres helper left at the default VOLATILE costs an index condition and a row estimate: a tenant holding 0.1% of 40,000,000 rows reads all of them.Thu, 29 Oct 2026 00:00:00 GMTdatabasespostgresperformancecontact@crimsonumbrella.com (Yunik Maharjan)A semantic cache breaks even at one repeat in 10,000 — then the threshold eats ithttps://crimsonumbrella.com/insights/semantic-caching-and-the-repeat-rate/https://crimsonumbrella.com/insights/semantic-caching-and-the-repeat-rate/Why a semantic cache breaks even at a 0.01 per cent repeat rate, how to set the threshold from a false-hit budget, and what the semantic half is worth.Thu, 29 Oct 2026 00:00:00 GMTcachingcostragcontact@crimsonumbrella.com (Saurav Raj Joshi)A 60-second auto-suspend bills 30% more a cycle than a 300-second onehttps://crimsonumbrella.com/insights/snowflake-auto-suspend-sixty-seconds/https://crimsonumbrella.com/insights/snowflake-auto-suspend-sixty-seconds/Two mechanisms get blamed for the Snowflake auto-suspend setting and only one of them is real. The crossover, derived from the gap between queries.Thu, 29 Oct 2026 00:00:00 GMTsnowflakecostwarehousecontact@crimsonumbrella.com (Saurav Raj Joshi)Spark's small-file crossover is 4 MB a file, not 40,000 fileshttps://crimsonumbrella.com/insights/spark-slow-because-of-forty-thousand-files/https://crimsonumbrella.com/insights/spark-slow-because-of-forty-thousand-files/Spark prices every file open at 4 MB of scanning, which puts the small-file crossover at 4 MB of mean size, about 40x before listing becomes the bottleneck.Thu, 29 Oct 2026 00:00:00 GMTsparklakehouseperformancecontact@crimsonumbrella.com (Saurav Raj Joshi)The catalog decides which engines can write, not the file formathttps://crimsonumbrella.com/insights/the-catalog-decides-not-the-file-format/https://crimsonumbrella.com/insights/the-catalog-decides-not-the-file-format/Iceberg and Delta both commit by swapping one metadata pointer. The catalog holds that pointer, so the catalog decides which engines can commit.Thu, 29 Oct 2026 00:00:00 GMTlakehousearchitecturewarehousecontact@crimsonumbrella.com (Saurav Raj Joshi)Doubling the chunk fixes 52% of broken tables and costs 19% of the signalhttps://crimsonumbrella.com/insights/the-chunk-that-lost-the-table-header/https://crimsonumbrella.com/insights/the-chunk-that-lost-the-table-header/Why the splitter loses the number and the model never had it: split probability derived at LlamaIndex's shipped defaults, compounded across a 1,200-table corpus, and what a bigger chunk costs.Thu, 29 Oct 2026 00:00:00 GMTragllmcontact@crimsonumbrella.com (Yunik Maharjan)PHP 8.1's CVE rate doubled after end of life. Its live siblings rose fasterhttps://crimsonumbrella.com/insights/the-end-of-life-cliff-is-a-date/https://crimsonumbrella.com/insights/the-end-of-life-cliff-is-a-date/PHP 8.1 reached end of life on 31 December 2025. Canonical patched 22 new CVEs into it in the 301 days after: 2.22 a month against 1.11 before.Thu, 29 Oct 2026 00:00:00 GMTsecurityvulnerabilitiescostcontact@crimsonumbrella.com (Yunik Maharjan)An output token takes 142x an input token, and the KV cache is whyhttps://crimsonumbrella.com/insights/the-kv-cache-is-the-bill/https://crimsonumbrella.com/insights/the-kv-cache-is-the-bill/Tokenisation to the KV cache, then the arithmetic: 128 KiB a token on Llama 3.1 8B, 16 GiB of cache for one full context window, and why prompt caching buys 22 per cent of the bill and 7 per cent of the clock.Thu, 29 Oct 2026 00:00:00 GMTllmcostcachingcontact@crimsonumbrella.com (Yunik Maharjan)Register as a company and your mobile ship tail drops from 28 days to 7https://crimsonumbrella.com/insights/the-mobile-bill-of-materials/https://crimsonumbrella.com/insights/the-mobile-bill-of-materials/What a mobile ship estimate leaves out: store accounts, signing identities, a push key, privacy declarations, and the wait your account type buys you.Thu, 29 Oct 2026 00:00:00 GMTmobilecostcontact@crimsonumbrella.com (Rujen Shakya)npm fixed install scripts in July, and 24 Node releases later yours still runs themhttps://crimsonumbrella.com/insights/the-post-install-script-is-the-attack-surface/https://crimsonumbrella.com/insights/the-post-install-script-is-the-attack-surface/npm 12 blocked dependency install scripts by default on 8 July 2026. No Node.js release bundles it, so the default on your machine still runs everything.Thu, 29 Oct 2026 00:00:00 GMTsupply-chainsecurityvulnerabilitiescontact@crimsonumbrella.com (Yunik Maharjan)Rotating the secret removes none of the six copieshttps://crimsonumbrella.com/insights/the-secret-in-the-spark-config/https://crimsonumbrella.com/insights/the-secret-in-the-spark-config/A client secret handed to a Spark job lands in six places. Rotation deletes none, and a derived 30-day period still runs 576 times a federated token.Thu, 29 Oct 2026 00:00:00 GMTsecuritysparkarchitecturecontact@crimsonumbrella.com (Yunik Maharjan)Token prices fell 73% and the same answer costs 2.7 times morehttps://crimsonumbrella.com/insights/token-prices-fell-and-your-bill-went-up/https://crimsonumbrella.com/insights/token-prices-fell-and-your-bill-went-up/List prices fell 73.3% from Claude Opus 4.1 to Opus 5.5, and the same answer costs 2.687x more as a ten-turn loop. The caching crossover is 59.6%.Thu, 29 Oct 2026 00:00:00 GMTcostagentsllmcontact@crimsonumbrella.com (Saurav Raj Joshi)A Tesla Roadster hovers on 28.3 kilograms of nitrogen a second, and 25.6 seconds is the ceilinghttps://crimsonumbrella.com/insights/twenty-eight-kilograms-a-second/https://crimsonumbrella.com/insights/twenty-eight-kilograms-a-second/Hover thrust equals weight, so 1,611 kg of car and bottles throws 28.3 kg of nitrogen a second. Fifty bottles buy 5.03 seconds, and no bottle buys 25.6.Thu, 29 Oct 2026 00:00:00 GMTphysicsperformancecostcontact@crimsonumbrella.com (Yunik Maharjan)UTC+05:45 is worth 15 minutes a day; New York costs 105 on ESThttps://crimsonumbrella.com/insights/utc-plus-five-forty-five-breaks-schedulers/https://crimsonumbrella.com/insights/utc-plus-five-forty-five-breaks-schedulers/Nepal Time is UTC+05:45. What the quarter-hour is worth against London and New York, what cron does with it, and why the working week costs more.Thu, 29 Oct 2026 00:00:00 GMTtimezonesorchestrationreliabilitycontact@crimsonumbrella.com (Rujen Shakya)Below $60,000 of decisions, the honest answer is to hire nobodyhttps://crimsonumbrella.com/insights/what-a-part-time-technical-lead-costs/https://crimsonumbrella.com/insights/what-a-part-time-technical-lead-costs/Four ways to buy technical leadership, costed over the same six months, and the $60,000 of decision exposure below which none of the paid ones pays for itself.Thu, 29 Oct 2026 00:00:00 GMTcosteconomicscareerscontact@crimsonumbrella.com (Saurav Raj Joshi)A six-hour streaming backlog replays in four hours, or in sixtyhttps://crimsonumbrella.com/insights/what-real-time-costs/https://crimsonumbrella.com/insights/what-real-time-costs/A worked replay calculation for a streaming pipeline, the latency floor exactly-once imposes, and why the first stream costs a rota rather than a processing job.Thu, 29 Oct 2026 00:00:00 GMTstreamingarchitecturecontact@crimsonumbrella.com (Saurav Raj Joshi)Your incremental model rewrites every file to change 0.95% of the rowshttps://crimsonumbrella.com/insights/when-a-full-refresh-beats-incremental/https://crimsonumbrella.com/insights/when-a-full-refresh-beats-incremental/A dbt merge scans the target and rewrites whole files, so the crossover isn't a ratio in rows: the rebuild budget puts it at 0.00053% churn.Thu, 29 Oct 2026 00:00:00 GMTwarehousecostpipelinescontact@crimsonumbrella.com (Saurav Raj Joshi)70 credits a day for 96 minutes of work: why a warehouse bill climbs on flat datahttps://crimsonumbrella.com/insights/why-the-warehouse-bill-climbs-on-flat-data/https://crimsonumbrella.com/insights/why-the-warehouse-bill-climbs-on-flat-data/Where the money goes when a Snowflake or Databricks bill rises on flat data: count the meters, attribute per query, check auto-suspend, then warehouse size.Thu, 29 Oct 2026 00:00:00 GMTcostwarehousecontact@crimsonumbrella.com (Saurav Raj Joshi)The fourth WordPress plugin costs 4x the obvious estimatehttps://crimsonumbrella.com/insights/wordpress-until-the-fourth-plugin/https://crimsonumbrella.com/insights/wordpress-until-the-fourth-plugin/Plugin maintenance is superlinear in the plugins that own a write path, and the crossover against a half-day-a-month budget lands at 3.45.Thu, 29 Oct 2026 00:00:00 GMTarchitecturecostdatabasescontact@crimsonumbrella.com (Sahil Shrestha)