88 KB of read-only rows priced Aurora Serverless v2 out of an agentic RAG rewrite

Aurora Serverless v2 does offer a 0-ACU floor, and its timers are what rule it out here. The minimum auto-pause interval is 300 seconds, a resume takes about 15 seconds, and after a day idle it is 30 or more ^[6]. A demo visited twice a week leaves gaps of roughly three and a half days, about a thousand times the shortest pause interval available, so every session starts from a paused cluster ^[2]. The author wrote that at that traffic, "30 seconds before the first byte" is the user experience ^[8]. Held at 0.5 ACU instead, the same traffic pays about $5 a session in idle capacity ^[1].

Then the inventory: 278 projects and a few thousand library rows, 88 KB, rebuilt from scratch by the ingestion pipeline and read-only at query time ^[9]. "That is not a database workload. That is a file," the author wrote ^[10]. The read path became a projects.sqlite artifact holding the projects and libraries tables, the corpus and an FTS5 index over it, built by the pipeline, published to S3 and pulled into the Lambda's /tmp on cold start ^[11]. Conversations, feedback and the spend ledger went to DynamoDB ^[12]. About $43 a month became about $1 ^[13], or $504 a year ^[3].

The quality argument runs the same direction as the price. Postgres full-text search has no IDF at all: ts_rank_cd scores on term frequency and position and is blind to the corpus, while SQLite's bm25() does weight rare terms ^[14]. The app was replacing minsearch indexes ^[2], so on this corpus the file ranks closer to what it displaced than the managed engine would have ^[14].

Two queries needed rewriting, not placeholder substitution. SQLite has no DISTINCT ON, and a GROUP BY with MAX() returns the row that produced the maximum, which is the row DISTINCT ON would have kept ^[17]. ARRAY_AGG(repo ORDER BY attempt) has no portable equivalent, because SQLite only gained ordered aggregates in 3.44 and the Lambda runtime may ship something older, so that grouping moved into Python ^[18]. The rest carried over because the project's own schema.sql opens with a note that these queries ran on SQLite before the Postgres migration and that the SQL had been kept portable ^[15]. "Past me left a door open," the author wrote ^[16].

Three conditions carry the $43-to-$1 swap to another stack: the query-time data is read-only, it is small enough for the pipeline to rebuild and ship as one artifact, and the writes fit a key-value access pattern ^[9][11][12]. A schema leaning on Postgres-only features costs more than two rewrites. The post gives the database line item and does not total the new stack's monthly bill ^[3].

Streaming went the same way. Lambda response streaming is supported on Node.js managed runtimes only, and Python needs a custom runtime or the Lambda Web Adapter ^[19]. The router runs up to six turns of tool calls plus a query rewrite and a rerank, and streaming only helps the final call, so the user waits through the tool chain either way ^[20]. A client that disconnects mid-stream is still billed the full function duration and the full token spend ^[21]. The app ships buffered JSON ^[22].

The last item is a signing detail. The plan was CloudFront with Origin Access Control in front of a Lambda Function URL on AuthType: AWS_IAM, with OAC signing the requests and no shared secrets ^[23]. CloudFront signs the request but does not compute the body payload hash, and Lambda rejects unsigned payloads, so /api/ask and /api/feedback both fail with InvalidSignatureException ^[24]. The architecture the post describes puts the Lambda behind a shared-secret header ^[3].