Rendered HTML Table Rows

Export the table you can see after a public page renders. Supply up to five page URLs and a CSS selector or table index. Each dataset item is one observed data row, with merged cells expanded, header paths preserved, and links and cell origins attached.

This is a bounded snapshot. It does not scroll, paginate, log in, solve challenges, or establish that every row in a remote application has loaded.

Start with one table

{
  "pageUrls": ["https://www.w3schools.com/html/html_tables.asp"],
  "tableSelector": "#customers",
  "maxRows": 10,
  "maxChargeUsd": 0.01
}

For this ten-row example, set the separate platform maximum charge to $0.06: the $0.05 start plus up to ten $0.001 rows. That option is a run control, not an extra Actor input property. Use 2048 MiB, a 210-second platform timeout and restart off.

Missing API input and {} use that public demonstration page, the #customers table, and a 100-row maximum. Set the platform maximum charge to $0.15 to fund all100 default rows plus the start. An explicitly empty URL list is invalid. When supplying your own URLs, select a table or leave tableSelector empty to use tableIndex (zero-based). Supplying URLs without a selector uses index 0. A nonempty selector and a nonzero index conflict and fail before navigation.

The page loads once. The Actor waits up to waitForTableMs for the selected visible table, then waits settleMs once before taking the snapshot. Increase those bounded delays for tables that load slowly. There is no claim that a quiet table has finished loading.

Read the result

The package includes sample_row.json, generated from an original synthetic table with a reserved example hostname.

values maps deterministic column keys to strings or null. Blank source cells remain ""; absent grid slots are null. Numbers and dates stay source text. Unicode is preserved and visible whitespace is collapsed. Header paths combine leading header rows, including multi-row thead sections. Empty names become column_1, etc.; duplicate keys receive deterministic numeric suffixes.

cells contains each logical column's text, visible links, header path and origin. Origin coordinates are zero-based: source row in the selected table, source cell within that row, and expanded grid column. A repeated span value retains the same origin, declared spans and effective row-group span. isSpanCopy identifies repeated values. rowspan="0" extends only to the end of its row group. Overlapping spans fail explicitly rather than shifting values.

Hidden rows do not become dataset items. Hidden cells retain their grid position with empty text. Nested table rows are not separate rows of the outer table; their visible text can appear inside the containing cell. Footer rows remain data rows. This extractor reads HTML table elements, not arbitrary ARIA or CSS grids.

Links include their source href, visible text and resolved HTTP(S) URL. Non-web or credential-bearing links have a null resolved URL. An overlong link is flagged and never resolved from a truncated prefix. Links are evidence only and are not followed.

record_id hashes the normalized input URL, table index and source row coordinate. It remains stable for the same position, but it is not an entity identifier when a source reorders rows. retrieved_at, final_url, and coverage: "OBSERVED_DOM_ONLY" retain source context.

Outcomes and delivery

The default dataset contains useful row records. The OUTPUT key-value record contains per-page outcomes, resource counters and a separate delivery receipt. A missing table with zero wait reports TABLE_MISSING; expiry with a positive wait reports TABLE_TIMEOUT. Empty tables, HTTP failures, unsupported sources and capped partial tables remain distinguishable. A bad page does not discard rows from healthy pages.

Run states are COMPLETE, PARTIAL, CAPPED, SOURCE_FAILED, INVALID_INPUT, DELIVERY_INCOMPLETE or RUNTIME_FAILED. COMPLETE means the bounded observation and delivery finished, not that the source application's full dataset was discovered. Common page warnings include ROW_LIMIT, REQUEST_LIMIT, DOM_BYTE_LIMIT, DOM_NODE_LIMIT, NETWORK_BYTE_LIMIT, OUTPUT_BYTE_LIMIT and field truncation limits.

A dataset append is attempted once. An uncertain response is recorded as unknown and fails the run; it is never silently retried. Confirmed delivery survives later bookkeeping errors. Completed same-storage restarts restore the row receipt without fetching or appending again. Incomplete restarts fail without replay. These replay paths have local regression evidence, not a hosted recovery guarantee. A new platform start can still incur the automatic start fee; restoring an old row receipt does not make the current start free. Keep restart off. Start a new run only after inspecting the original dataset when delivery is unknown.

Billing contract

The required billing configuration has two automatic events and no other positive event: apify-actor-start at $0.025 per allocated GB unit, and apify-default-dataset-item at $0.001 per delivered row. The fixed 2048 MiB allocation corresponds to two start units, totaling $0.05 per run. The platform charges the start automatically; the Actor never charges it manually.

When a run actually starts, the start fee can apply even if the page has no table, the source fails, runtime input validation fails, the local row budget is zero, or a later execution restores a saved row receipt. A platform input-schema rejection before any run exists is different and does not establish a start event. The fee pays for starting the bounded work, not for a promised result. No row is charged unless it is delivered to the default dataset. A six-row result therefore has a configured fee of $0.056, not $0.006. Actual platform receipts determine charged amounts; these examples are not statements that an owner test or unpriced run paid that fee.

maxChargeUsd remains a row-selection budget. It floors to whole candidate rows; below $0.001, no page is visited. It excludes the start fee. Do not put the overall run limit into this field. The separate platform maximum needs $0.05 plus the intended row budget, and can limit actual delivery below the selected row count. Use at least $0.051 for the separate platform maximum; setting a local row budget of zero is not a promise of a zero-cost platform run.

Example input file	Local row budget	Separate platform maximum
examples/defaults.json or {}	$0.10 /100 rows	$0.15
examples/selector.json	$0.01 /10 rows	$0.06
examples/index.json	$0.01 /10 rows	$0.06
examples/zero-cap.json	$0 /0 rows, no navigation	$0.051 minimum; start may still cost $0.05

billing.chargedItems counts only dataset-row events. On replay it remains the saved row receipt. startBilling describes the current execution separately: its expected tariff and units are configuration facts, while actualChargedUnits and actualChargeUsd remain null because this Actor does not independently observe the platform's start-charge receipt. PPE_VALIDATED confirms the exact configured prices, not a charge; UNPRICED or NOT_CHECKED is not a paid receipt. Owner-accounted zero events are QA, not customer revenue. Platform compute, transfer, storage and later inspection are separate economic liabilities; this local budget is not a lifetime cost bound.

Hard limits

Boundary	Default	Maximum
URLs	1 example	5 supplied
Rows across pages	100	1,000
Columns	64	64
Table appearance wait	1,500 ms	5,000 ms
Extra settle delay	300 ms	2,000 ms
Page work	15 seconds	20 seconds, plus bounded cleanup
Source collection	120 seconds	120 seconds
Whole workflow	180 seconds	plus up to 4 seconds for timeout receipts
Admitted HTTP requests/page	40	80, including redirects
Main-frame navigations/page	6	6
Relayed socket bytes/page	4 MiB	8 MiB
Observed DOM bytes/page	512 KiB	2 MiB
Observed DOM nodes/page	10,000	20,000
Cell text	4,096 characters	4,096
Links/cell	16	16
Serialized output	8 MiB	8 MiB and one SDK append chunk
Browser process-tree RSS	1 GiB	sampled every 50 ms

DOM bytes count node names, text and attributes, not a serialized HTML document. Socket bytes include both directions and TLS ciphertext; rejected over-limit chunks are not relayed. Browser memory can briefly overshoot between samples. Use the required 2 GiB hosted container and a 210-second platform timeout for an additional process boundary.

Each page gets fresh cookies and storage. The browser receives no Actor token or stored profile. Public-IP checks apply to every upstream connection, including redirects and subresources, and resolved addresses are pinned to the socket. Only standard HTTP/HTTPS ports and GET requests are supported. Images, fonts, media, downloads, service workers, frames, popups, worker targets and WebSockets are blocked. A context-wide request gate is installed before page creation to intercept a popup's first request; late window closing alone is not the exclusion mechanism. Restrictions can prevent a page's table from rendering; the Actor does not bypass them. Browser-supported gzip, deflate and HTTPS Brotli are accepted; damaged encodings fail as source navigation errors.

Per-page childRequestGateAttempts and childRequestGateCounts record child-target requests intercepted before attempting an abort, using only method/type labels. These attempts also count in requestAttempts and blockedRequests. childRequestAbortsConfirmed counts browser acknowledgments of aborts, not a remote server observation. A gate count is not a count of created windows: workers or frames can also reach it, while other restrictions can prevent them earlier. No child URLs, headers or credentials are logged in these counters.

Known login and challenge markers and declared virtual row counts are rejected. Such detection is incomplete: all other results still carry observed-DOM-only coverage. Use only public pages you are authorized to access.