Build a scheduled scraper pipeline that's safe to run unattended for months

Build a scheduled data pipeline that runs once every night at 2am UTC, pulls a public list of competitor product prices from 40 known product URLs, and stores clean, de-duplicated results in a database — built to run unattended for months, not just to work once when I watch it run manually.

PIPELINE STAGES
1. Fetch: pull from a public list of competitor product prices from 40 known product URLs. Respect a fixed 2-second delay between requests, and treat a 429 or 503 response as a signal to back off and retry the whole run later rather than hammering the source immediately — this is someone else's server, and a scraper with no delay between requests is the kind of thing that gets an IP blocked, which would break every future run, not just this one.
2. Parse: extract product name, current price, currency, and in-stock status from the raw response.
3. Normalize: clean and standardize the extracted fields — strip currency symbols and commas from price into a plain decimal, normalize in-stock text ("In Stock", "Available", "2 left") into a single boolean is_in_stock.
4. Store: write to a database, keyed so that re-running the pipeline against the same source data never creates duplicate rows — use a natural unique key from the source data itself (not an auto-incrementing ID that would happily insert the same logical record twice) with an upsert, not a plain insert.

RELIABILITY RULES
If one item in a batch fails to parse or fetch, log that specific failure with enough detail to debug later and continue processing the rest of the batch — one bad item must never abort the whole run and silently skip every item after it. Store the raw fetched response alongside the normalized result, not just the normalized version, so that if a normalization rule turns out to be wrong later, the pipeline can be corrected and re-run against the original raw data rather than needing to re-scrape from scratch. Add a run log table recording every pipeline run — start time, end time, items processed, items failed, and a short status — so I can see the actual history of runs without digging through console output.

SCHEDULING AND ALERTING
Set this up using Replit's Scheduled Deployments (or the closest equivalent — cron-style scheduling, not a script that only runs while I happen to have the workspace open) to run once every night at 2am UTC. If a run fails outright — the source is unreachable, an unexpected exception escapes the per-item error handling above — send a notification to an email to me via a simple SMTP send, since this doesn't warrant a full Slack integration for one person so I find out the same day, not the next time I happen to check manually and notice weeks of missing data.

CONSTRAINTS
Replit's managed Postgres database, with a raw_snapshots table and a separate normalized products_daily table Handle the source being temporarily unreachable (a timeout, a 503) as a retryable condition with backoff, distinct from a genuinely malformed response that indicates the source's structure changed and needs a human to look at the parser, not just retry.