eroski¶

supermercapy.eroski.Eroski reads eroski's online supermarket at supermercado.eroski.es. the storefront runs apache tapestry 5 and renders every page on the server. it publishes no json api, so the client reads html: listing tiles, product pages and the navigation menu. caprabo runs the same build on its own host, and both clients share one parser.

nothing binds. the storefront pins the anonymous session to shop 157 (named bilbondo), which sets the prices and the regional range. choosing another shop needs an account, because the delivery pages redirect an anonymous visitor to a login, so store_id is None and each product carries the shop_id that priced it.

from supermercapy import Eroski

with Eroski() as eroski:
    page = eroski.search_products("aceite de oliva")  # one request
    for row in page.products:
        row.price.amount, row.price.reference, row.promotions

    oil = eroski.get_product(page.products[0].id)  # one request
    oil.nutrition.values[0].name  # "Energía"
    oil.manufacturer, oil.storage

    tree = eroski.get_categories()  # one request, about a megabyte
    fruit = eroski.get_category("2059699")  # one more: the menu is kept

capabilities¶

capability	supported	how
CATALOG	yes	iter_catalog reads the menu, then walks each root category's listing to its empty page
NUTRITION	yes	ingredients and the nutrition table from the product page's characteristics
PROMOTIONS	yes	multi-buy badges such as "3x2" and "2ª unidad -70%" become Promotion records
POSTAL_CODE / STORES	no	the shop follows the account; an anonymous session cannot pick one
OFFERS	no	/es/supermercado/filter/offers/ redirects an anonymous visitor to a login
NEW_ARRIVALS	no	nothing marks a product new
HOME	no	the home page carries banners and no product tiles
EAN / EAN_LOOKUP	no	no barcode appears in any page
FUTURE_PRICES	no	prices carry no dates

extensions¶

method	what it does
get_category_products(category_id, cursor=)	one listing page of a category, its subtree included, paged like search

the product model adds shop_id, is_marketplace for third-party sellers, is_lowered_price for the lowered-price badge, and, from a product page, manufacturer, manufacturer_address, alcohol_percentage and features, every characteristics box as a (title, text) pair. categories add path, the slug path a listing is addressed by, and url.

request counts¶

call	requests	size
search_products	one per page	about 230 kb
get_product	one	about 900 kb
get_categories	one	about a megabyte
get_category, get_category_products	one, plus the menu once per client	about 230 kb
iter_catalog	the menu, then one per page of each root plus its empty page	hundreds of pages

the sizes are the html after decompression. the storefront compresses its answers, so the wire carries about a tenth of that: twelve requests that decompress to 4.5 mb arrived as 330 kb in october 2026.

quirks¶

listings come from the infinite scroll's fragments. a full listing page weighs a megabyte, most of it the menu. the client asks for the zone fragment the storefront's own scroll loads instead, …:loadpage?…&t:zoneid=productListZone&pageNumber=N, which answers json holding about twenty tiles as html for about 230 kb. a page past the end answers with no tiles, and that empty page ends a walk. a one-page category says it is finished on its first page, so it costs no empty page.

pages hold about twenty rows and take no size. max_page_size is 20. the client cuts a smaller page_size out of one storefront page, so the cursor names the storefront page and the rows of it already returned, "3" or "3:5", and walking a small page size asks for the same storefront page again until its rows run out. a root category's first page sometimes holds 21. the storefront publishes no total, so total_hits is always None.

the menu is the category tree. every full page carries the whole navigation, about 1,500 nodes four levels deep, and each link's path spells the node's ancestry: /es/supermercado/2059698-frescos/2059699-frutas/. the client rebuilds the tree from those paths. a few nodes sit under two parents in the menu, and the tree keeps both placements, with the same id and a different path. the menu includes eroski's non-food roots, such as electrónica and electrohogar.

a listing needs the full slug path. a fragment addressed by a bare category id answers empty, and one addressed by the leaf's own segment left out a product the full path listed in october 2026. so get_category reads the menu first, once per client, and the client keeps the tree for later calls. get_categories always fetches a fresh one. an id the menu does not hold raises NotFoundError. no listing publishes a product count, so product_count is None.

ids are opaque strings. most are numeric, and marketplace products read MP96290. the storefront ignores the slug in a product url, so the client asks for /es/productdetail/<id>-producto/. an id with no slug at all, or an unknown one, redirects to /es/error/404/, which raises NotFoundError. product pages name no canonical url, so a product from get_product has no slug.

errors inside fragments are redirects in json. an event the storefront cannot serve answers http 200 with {"_tapestry": {"redirectURL": …}} instead of content. a redirect to the 404 page raises NotFoundError, and any other, such as the general error page a blank query gets, raises InvalidResponseError. the client refuses a blank query before sending it.

the visible price wins over the analytics one. each tile carries a ga4 data-metrics attribute with the id, name, brand, price and category ids. its price is a float that drops trailing zeros, so price.amount reads the price printed on the tile, and falls back to the analytics figure when the tile prints none. a struck-through price fills previous, and the percentage badge beside it fills discount_percentage.

units. the unit price reads 1 LITRO A 0,97 € in every language. KILO, LITRO, DOSIS, DOCENA, UNIDAD and METRO are in the shared vocabulary, and ROLLO reads as one piece, so price.reference restates the figure per kilogram, litre, dose, piece or metre. most single-unit products print no unit price at all, and then reference is None.

purchase limits come from a script. the page's tapestry init script builds each add-to-cart button from options: the most a basket may hold, the smallest quantity, the pack size, whether the product is sold by weight, the seller and the shop. they fill availability and shop_id. produce sold by weight also shows a weight icon, and either source sets is_variable_weight.

some products are filed nowhere. a product page's breadcrumb sometimes stops at the home link, and the analytics event then names no category, so category_path and category_ids stay empty for those products.

the client reads the nutrition table by position. its first row states what the figures refer to, such as 100 mililitros, and fills nutrition.per. figures against a hundred grams or millilitres land in per_100, anything else in per_serving. the manufacturer box alternates a bold label with its value, and the labels follow the language, so the first value is manufacturer and the second manufacturer_address.

the user agent. the edge refuses a few user agents, among them curl, python-requests and an empty one, with a bare 134-byte http 403. the edge accepts the library's own user agent, so it stays the default. a 403 raises BlockedError after one request.

datacenter addresses get a recaptcha page. google cloud armor answers a client it doubts with its "checking your browser" recaptcha page and http 200, in place of a page, a json fragment or a photo. on 4 october 2026 a hetzner server got it on every request, at eroski and caprabo alike, while a residential connection never did. the client recognises the page and raises ChallengedError after one request rather than trying to parse it, and it never tries to solve the challenge.

pacing. min_request_interval defaults to one second, because a catalog walk is hundreds of quarter-megabyte pages. eroski showed no rate limit or challenge in october 2026.

languages. the first path segment picks the language, /es/, /ca/ or /en/. the menu and the page labels follow it, and product names stay spanish. the unit price text stays spanish in every language. eroski also speaks basque, galician and german, which Language does not hold.

robots.txt. eroski's robots.txt disallows /*?t:ac=, /*?filter=, */search/results/?q=, */?zipCode, the language switch, a chatbot event and session ids. the client sends category listings to /es/supermarket:loadpage?t:ac=…, which the /*?t:ac= rule covers. searches go to /es/search/results:loadpage?q=…, which does not match the search rule's path literally but is the same search. the owner chose to use these paths.

non-goals¶

• choosing a shop. the anonymous session cannot, and ?zipCode= changes nothing. logging in to pick a shop is out of scope.
• offers and filters. the offers page needs a login, and the filter menu stores its state in the session through tapestry events. the client never changes session state.
• reviews. product pages carry customer reviews and ratings. the client does not parse them.