Under the hood

How the platform is operated

Software a restaurant runs service on fails at the worst possible time: seven on a Saturday. What matters is whether someone knows within minutes, or hears about it from the restaurant on Monday.

26 statements, each traced to code or a test.

All technical pages

What a restaurant cannot see from a feature list

Two products can list the same features and behave very differently on a bad night. The difference is in the parts no screenshot shows: whether an error is noticed, whether a change to the database was rehearsed before it touched live bookings, whether the checks that run before a release would have caught the fault.

For the restaurant the consequences are specific. Bookings that silently stop arriving. A payment that is taken but not recorded. An update that works in a demo and breaks on the one restaurant with an unusual menu. Each of these costs covers and trust, and none can be put right from the restaurant’s side of the screen.

This page sets out how TableSpark is run day to day, with the mechanisms named so they can be checked.

In short
  • Every request and every scheduled job in the service runs inside a monitor that records uncaught errors and server-error responses.
  • A separate monitor checks five things once a minute: the service’s health address, a restaurant website, the tablespark.uk home page, the sign-in configuration and the database’s sign-in service.
  • On 11 November 2026 a batch of 47 migrations was applied first to a disposable, production-shaped Postgres 17 database, with each file run twice, before it was applied to production.
  • The repository holds more than 800 automated test files.

Errors are noticed

Every request and every scheduled job in the service runs inside a monitor that records uncaught errors and server-error responses. Monitoring is switched on in the production configuration.

The monitor also collects errors from the dashboard in the browser, scheduled jobs that fail or go missing, and queues that stop moving: campaign email, guest messages, journeys, push notifications, print jobs and payment reconciliation alerts. Error samples are scrubbed of secrets, tokens, contact details and card numbers before they are stored.

UrgentProblems on booking, ordering, payment and sign-in routes are top priority. They raise an alert at once, repeat it at 30, 60 and 120 minutes until someone acknowledges it, and report recovery after 15 quiet minutes.
Everything elseLower-priority problems arrive as an hourly digest or a daily summary, so the urgent alerts stay rare enough to be read.
Scheduled jobsEach scheduled job writes a heartbeat when it finishes, and a job whose heartbeat is older than twice its interval plus five minutes is flagged as late.
One viewThe platform team works from a Health view with a tile for each surface (sites, bookings, orders and payments, email, the Guest App, scheduled jobs and AI), the open problems, and the time each scheduled job last finished.

A public address, /_ts/ops/pulse, reports whether monitoring is on and how many scheduled jobs are late.

A monitor that stands apart

A system cannot be relied on to report its own outage. The checks behind the public status page therefore run somewhere else.

A separate monitor checks five things once a minute: the service’s health address, a restaurant website, the tablespark.uk home page, the sign-in configuration and the database’s sign-in service. It runs apart from the systems it watches and keeps its own records, so it goes on measuring when the main service or the database is down.

An interruption is recorded after two failed checks in a row and closed after two passes in a row. The results are public on the status page and as JSON at /_status/summary, with 90 days of daily history.

The current state is on the status page.

Scheduled work

Scheduled work runs on six timers: one every minute for email delivery, four that each fire every quarter of an hour at staggered minutes, and one daily run. The daily run applies the data-retention rules and re-checks the refund ledger.

Changing the database safely

The riskiest change to a booking and payments system is a change to its database, because a mistake there is applied to every restaurant at once.

The database is built by more than 350 migration files, applied in order.

  1. Rehearsed firstOn 11 November 2026 a batch of 47 migrations was applied first to a disposable, production-shaped Postgres 17 database, with each file run twice, before it was applied to production.
  2. Compared afterwardsAfter that release the production catalogue of functions, policies, columns, triggers and indexes was compared with the simulation, object by object.
  3. Permissions gatedA gate test reads the real permissions of every privileged database function and fails if one is reachable by the public outside a fixed allowlist, or is missing a pinned search path. The reviewed inventory lists more than 950 such functions.

Automated tests

The repository holds more than 800 automated test files. More than 120 of them run against a real Postgres database that is started for the test, loaded with the full migration chain and thrown away afterwards. They never touch production data.

A further suite runs the real service code in the Workers runtime against that database, including the payment-notice and order-state cases.

The tests cover bookings and tables, online ordering and payments, the till, gift cards and stock, the Guest App, loyalty and email, the site builder and published sites, sign-in and data protection, and the menu scan.

The site build stops if a canonical address, a sitemap entry or the robots file names a host other than production, or if a shipped page imports a module that is missing from the build.

Certified, monitored and insured

TableSpark Ltd is Cyber Essentials certified for the whole organisation under the UK Government-backed scheme run by the National Cyber Security Centre with IASME. The certificate was issued on 17 November 2026, is valid to 17 November 2027 and can be checked on the public registry.

Our domain is enrolled in the National Cyber Security Centre’s Early Warning service, which alerts us to potential threats and vulnerabilities it detects on our network. We also carry cyber liability insurance, underwritten by AIG UK, which brings incident response specialists in from the first call. The security page sets out what each covers.

What it runs on

The service runs on Cloudflare Workers and Pages, with the database, sign-in and file storage on Supabase.

The changelog lists what has shipped, with dates, and the security page covers payments, sign-in and data protection.

Questions owners ask

How does TableSpark find out that something has broken?

Every request and every scheduled job in the service runs inside a monitor that records uncaught errors and server-error responses. Problems on booking, ordering, payment and sign-in routes are top priority. They raise an alert at once, repeat it at 30, 60 and 120 minutes until someone acknowledges it, and report recovery after 15 quiet minutes.

Is the status page checked by the same system it reports on?

A separate monitor checks five things once a minute: the service’s health address, a restaurant website, the tablespark.uk home page, the sign-in configuration and the database’s sign-in service. It runs apart from the systems it watches and keeps its own records, so it goes on measuring when the main service or the database is down.

Are database changes tested before they reach live restaurants?

On 11 November 2026 a batch of 47 migrations was applied first to a disposable, production-shaped Postgres 17 database, with each file run twice, before it was applied to production. More than 120 of them run against a real Postgres database that is started for the test, loaded with the full migration chain and thrown away afterwards. They never touch production data.

How this page is kept true

This page makes 26 statements about how TableSpark behaves. Each one is recorded with the place in the code, the database or the test suite that enforces it, and an automated check fails when a statement loses its evidence. The suites named below exercise the behaviour described here.

Automated suites behind this page

  • db-audit-chain-db
  • email-depth-db
  • gift-card-redemption-db
  • kitchen-display-db
  • menu-stock-db
  • ops-monitoring
  • ops-monitoring-wiring
  • reservations-tables
  • rpc-acl-gate
  • seo-tenant
  • site-redirects-db
  • status-monitor
  • terminal-ledger-db

See it on your own restaurant

Build your site free and publish when you are ready. Plans start at £19 a month, excluding VAT.

Start building free