internal/eusupply replaces the old Node scraper (eu-supply.js) for
uk.eu-supply.com. The site has no API, so it posts the public search
form, which now needs the form's current defaults and reads the HTML
with golang.org/x/net/html. Each tender not stored yet has its own
page read too, for its description, contract type, CPV codes and
buyer address. Tenders are stored with the source "EU Supply - UK"
as eusupply-uk-<tender id>, with one release holding the tender as
read.
The site publishes no rate limit, so the client sets its own: one
request at a time, with a random wait between
sources.eu_supply.request_min_gap and request_max_gap (2s to 10s).
The source has the same 7-day backfill as Find a Tender's.
Most of its tenders are also on Find a Tender, so a new one is stored
Hidden and Skipped when another source's opportunity released within
30 days has a title at least 90% alike, and its buyer or description
matches too (80% of the shorter one's significant words).
Backfilling a long window (an early --since, or a long time offline)
hit the API's rate limit, which could still be exceeded after waiting
once, and then everything fetched so far was lost.
Before each Find a Tender source's worker starts, a start point more
than 7 days ago is now backfilled in 7-day windows bounded on both
sides, oldest first, each saved before the next is fetched. A window
that fails is saved as far as it got and fetched again from its start
after the source's interval. The worker then carries on from the end
of the last window.
A 429 is now waited out up to sources.<source>.rate_limit_retries
times (default 5) rather than once; fts-once --ignore-rate-limit uses
0. FetchAll returns what it fetched before an error along with it, and
the workers save it, so a failed run no longer loses its data.
findatender's FetchAll and FetchOnce take RequestOptions, which gains
Until (updatedTo), and RateLimitRetries replaces AllowRateLimit.
sources.skip_rules_file names a JSON file of filters, each with a
comment, an active flag, a logic mode ("or" by default, or "and") and
rules by buyer, title words or CPV codes. The fetcher applies them
before saving, so only new opportunities are marked; the marks of
stored ones are never touched. CPV rules match when the opportunity's
codes are all in the list, or any of them with exclusive false, and
by prefix with exact false.
The file is read again when its time or size changes: an invalid edit
is logged and the last good rules kept, and a deleted file means no
rules until it's back. fetcher service and serve refuse to start if it
is define but missing or invalid.
job-search skip-rules check validates a rules file, and skip-rules
apply marks the stored opportunities that match Skipped, or only lists
them with --dry-run.
Every save that stores new opportunities or new releases records them in a
change batch (migration 00008: opportunity_change_batches and
opportunity_change_batch_opportunities; the opportunities table is
unchanged), as does every change of marks. A batch is named by a random
SHA-256 hash, so it can't be guessed, and ordered by its created_at, which
is kept strictly increasing even if the clock goes back.
GET /api/opportunities pages now carry the newest batch, read before the
page, and GET /api/opportunities?changed_after=<batch> returns every
opportunity changed since that batch, unpaged, with the newest batch it
includes. Offset and limit can't be combined with it, and an unknown batch
is 410 Gone, telling the client to reload the whole list.
OpportunitiesUpdated now carries its save's batch, and a "large" flag: over
50 opportunities, new and updated are left empty and only the summary
counts them, so a big catch-up no longer sends tens of thousands of UUIDs.
SaveOpportunities now reports the classification codes a save brings into
use (AddedClassifications, in full) and takes out of use
(RemovedClassificationIDs), by comparing the codes in use before and after
the save inside its transaction. The fetcher publishes them as a
ClassificationsUpdated event, {"added":[codes],"removed":[ids]}, never null,
before the save's OpportunitiesUpdated, so clients can keep their list of
classifications without asking the API again.
internal/events decouples the parts of JobSearch that cause events from
the parts that react to them. bus.go has Bus: Subscribe, and a synchronous
Publish that calls each handler in turn on the publisher's goroutine, so
handlers mustn't block (the hub doesn't: it queues per client). events.go
has the event types and their payloads, starting with OpportunitiesUpdated.
With fetcher.WithEvents, a save that stores new opportunities, or new
releases of stored ones, publishes OpportunitiesUpdated with their UUIDs and
counts. An opportunity whose row is only rewritten, as when it is fetched
again from the newest stored release, isn't counted as updated, and nothing
is published when nothing changed. store.SaveResult now reports
NewOpportunity for each opportunity saved.
job-search serve subscribes its WebSocket hub to the bus, so every client
gets {"type":"event","eventType":"OpportunitiesUpdated","payload":{"new":
[...],"updated":[...],"summary":{"new":n,"updated":n}}}. The Envelope gains
an eventType field, and Hub.PushEvent and Server.PushEvent send events.
Without --since, fetcher service and job-search serve start each source
from the newest release stored for it (the new store.LatestReleaseDate,
by Source), so a restart carries on where it left off instead of fetching
the last day again; with nothing stored, the client's default still
applies. --since overrides it for every source. The Find a Tender source's
name is now a constant, as the Source its opportunities are stored with.
The README's guide to adding a source is brought up to date with the
current worker, start and config type names.
With WithStore, the Find a Tender and mock workers save each run's
opportunities with SaveOpportunities and log what changed; a failed save
fails the run, so the checkpoint stays put and the next run retries.
Without a store nothing is persisted. The mock's opportunities are saved
with Source "FTSMock" rather than the converter's "FindATender".
fetcher service and job-search serve open the store at database.file
(cli.OpenConfiguredStore), or warn and run without one. A missing file
isn't created, and a database needing migration says to run job-search
migrate.
Fixed upsertOpportunity, which never wrote the source column, so every
saved opportunity had an empty Source.
fetcher.Run(ctx, conf, since) becomes New(&conf.Sources, opts...) and
Run(ctx, since). Each worker now takes only its own source's config, and
Run starts each one through a closure that binds it. Cancelling ctx stops
the workers through the Fetcher's background context, and Run always waits
for every worker before returning their joined errors.
The findatender client no longer holds the configuration: New takes only
the API root, ReloadConfig is gone, and request tracing follows
logging.IsDebugEnabled(). The FindAOpportunity worker is now FindATender.
job-search serve runs the fetcher alongside the HTTP server. A source that
fails stops the server, and the server stopping stops the sources; with no
sources enabled the server keeps running. fetcher service and fetcher
fts-once use the new APIs.
An enabled Find a Tender or mock source with no api_root used to start
and then fail on every fetch. Its worker now returns an error naming the
missing key before it starts, which stops Run. "fetch once" checks
sources.find_a_tender.api_root the same way. TestRunMissingAPIRoot covers
both sources, and TestRunLogsSourceErrors sets an api_root so it still
tests the zero interval. The section docs and internal/config/README.md
say what an empty api_root means.
CODE_CONVENTIONS.md sets out the conventions for Go code in this
repository: naming, layout, comments, errors, logging, configuration and
tests. .golangci.json enables the linters that check the parts that can be
checked (each limited to the rules the conventions use); the rest is
applied by hand.
Applied across the codebase:
- Naming: full-word names in application code, initialisms in one case
(BuyerID, PrintJSON, PrintYAML, ocdsGetBuyerID), no names shadowing
built-ins (min/max are minGap/maxGap), mustRaw renamed rawJSON as it
returns nil rather than panicking, and config.ConfigLevelToLogLevel
renamed LevelToLogLevel.
- Layout: lines kept to 80 characters where possible, assignments taken
out of if and for statements that don't fit, wrapped arguments all on
one line or one per line, steps separated by blank lines, no else after
a return, function and struct literals wrapped.
- Comments: doc comments on every exported name (including each section's
Map, Flatten, TypeOf, ValueOf and SetValue), section banners, British
spelling.
- Errors: a created error starts with its package name, a wrapping error
adds only what was being done, sentinels in one commented var block,
errors.Is and errors.As in place of == and type assertions, and no
err := shadowing of a named result. That fixes the mock server losing
its access log's close error.
- Logging: the fetcher passes its workers an untagged logger, so
findatender's lines are tagged with their own component rather than
with fetcher's as well; cmd/mock-fts logs "stopped" with slog.
- Configuration: util.WithTimeout (moved from the websocket package)
applies a timeout of 0 or less as no limit, used for the websocket
write and ping, TCP close and tcpserver shutdown timeouts; every setting
documents what 0 or empty means, including the rate limit rule's keys
in internal/config/README.md.
- Tests: related tests merged into table-driven ones with named cases.
Separately from CODE_CONVENTIONS.md, internal/store/README.md gains a soft
SQL style convention for the queries in internal/store/queries.go, and the
queries are reformatted to follow it.
Short and single-letter names in application code become full words (tests
keep theirs). Wrapped calls, signatures and composite literals put one item
per line, and one-line function bodies are expanded. Exported names and
groups of constants get doc comments, and comments refer to the renamed
identifiers.
Log messages are lower-case event names with their details as key/value
attributes ("error" for errors), in place of formatted strings.
Loggable package take a WithLogger option as part of their "New*" call,
defaulting to slog.Default(). Each tags its lines with its `component`.
Some debug loggers print directly rather than using slog when in
terminal logging mode.
- Fixed function name typo
- Fetcher/Run now returns a combination of errors if any workers error out. All other workers are ordered to shutdown if any one does error out
- findATender now returns a pointer instead of an instance from New
Fetching:
- Add `fetch service`, which runs a worker per enabled source until
SIGINT/SIGTERM (internal/fetcher)
- Add util.Worker: runs a task on an interval with a since-checkpoint
that only advances on success, and a shutdown timeout for in-flight
work; plus AwaitTask/AwaitTaskE for calls that ignore cancellation
- Add fts_mock and sleeper sources
- Sleeper: sleeps periodically to emulate work being done. Can be
instructed to block the shutdown signal or handle it cleanly
Mock FTS server (cmd/mock-fts, internal/mock/findATender):
- Serves GET /api/1.0/ocdsReleasePackages with FTS paging, query
validation, UK local time handling and per-client rate limiting
- Generates fictional releases from an embedded catalog, moving
procedures through planning, tender and award, and keeps them in a
JSON Lines store that survives restarts
- Add logging.AccessLog for combined/common access logs
bwilliams