* persist queued worker tasks to database on shutdown, fill worker queues from database on startup
* ensure the tasks are sorted by creation time before pushing them
* add migration to insert WorkerTask{} into database, add test for worker task persistence
* add test for recovering worker queues from database
* quick tweak
* whoops we ended up with double cleaner job scheduling
* insert each task separately, because bun is throwing some reflection error??
* add specific checking of cancelled worker contexts
* add http request signing to deliveries recovered from database
* add test for outgoing public key ID being correctly set on delivery
* replace select with Queue.PopCtx()
* get rid of loop now we don't use it
* remove field now we don't use it
* ensure that signing func is set
* header values weren't being copied over 🤦
* use ptr for httpclient.Request in delivery
* move worker queue filling to later in server init process
* fix rebase issues
* make logging less shouty
* use slices.Delete() instead of copying / reslicing
* have database return tasks in ascending order instead of sorting them
* add a 1 minute timeout to persisting worker queues
* [feature] Federate interaction policies + Accepts; enforce policies
* use Acceptable type
* fix index
* remove appendIRIStrs
* add GetAccept federatingdb function
* lock on object IRI
* initial work replacing our media decoding / encoding pipeline with ffprobe + ffmpeg
* specify the video codec to use when generating static image from emoji
* update go-storage library (fixes incompatibility after updating go-iotools)
* maintain image aspect ratio when generating a thumbnail for it
* update readme to show go-ffmpreg
* fix a bunch of media tests, move filesize checking to callers of media manager for more flexibility
* remove extra debug from error message
* fix up incorrect function signatures
* update PutFile to just use regular file copy, as changes are file is on separate partition
* fix remaining tests, remove some unneeded tests now we're working with ffmpeg/ffprobe
* update more tests, add more code comments
* add utilities to generate processed emoji / media outputs
* fix remaining tests
* add test for opus media file, add license header to utility cmds
* limit the number of concurrently available ffmpeg / ffprobe instances
* reduce number of instances
* further reduce number of instances
* fix envparsing test with configuration variables
* update docs and configuration with new media-{local,remote}-max-size variables
* improved server shutdown with more precise shutdown of modules + deferring of ALL of it
* move delivery and workers into separate files
* add worker task model and Serialize() / Deserialize() methods for message types
* start adding message serialize / deserialize tests
* start adding test cases
* update body rewinding to rely on standard library mechanism of r.GetBody()
* remove request rewinding (http.Client{} should already handle this)
* standard library already handles rewinding
* improved code comment
* move the newPOST() function contents to prepare(), fits better with current API
* add Serialize() / Deserialize() implementations for Delivery{} type
* finish writing FromClientAPI sserialize / deserialize tests
* start adding FromFediAPI{} serialize / deserialize test cases
* fix FromFediAPI{} tests
* add tests for delivery message type
* fix repeat code
* missing license header
* use testrig status and accounts for marshal / unmarshaling tests
* add a specific test for checking account RSA keys are preserved
On startup and shutdown of a worker, we log a message of the worker
being started together with a textual representation of a memory
address. Though this can be handy for developers to debug
startup/shutdown sequencing issues of the workers, it's typically not
very useful or informative for an admin. We can also output a lot of
these (on my system I get 265 lines of these during startup).
This changes the messages from Info to Debug, to not print them under
normal circumstances.
* start replacing client + federator + media workers with new worker + queue types
* refactor federatingDB.Delete(), drop queued messages when deleting account / status
* move all queue purging to the processor workers
* undo toolchain updates
* code comments, ensure dereferencer worker pool gets started
* update gruf libraries in readme
* start the job scheduler separately to the worker pools
* reshuffle ordering or server.go + remove duplicate worker start / stop
* update go-list version
* fix vendoring
* move queue invalidation to before wipeing / deletion, to ensure queued work not dropped
* add logging to worker processing functions in testrig, don't start workers in unexpected places
* update go-structr to add (+then rely on) QueueCtx{} type
* ensure more worker pools get started properly in tests
* fix remaining broken tests relying on worker queue logic
* fix account test suite queue popping logic, ensure noop workers do not pull from queue
* move back accidentally shuffled account deletion order
* ensure error (non nil!!) gets passed in refactored federatingDB{}.Delete()
* silently drop deletes from accounts not permitted to
* don't warn log on forwarded deletes
* make if else clauses easier to parse
* use getFederatorMsg()
* improved code comment
* improved code comment re: requesting account delete checks
* remove boolean result from worker start / stop since false = already running or already stopped
* remove optional passed-in http.client
* remove worker starting from the admin CLI commands (we don't need to handle side-effects)
* update prune cli to start scheduler but not all of the workers
* fix rebase issues
* remove redundant return statements
* i'm sorry sir linter
* add delivery worker type that pulls from queue to httpclient package
* finish up some code commenting, bodge a vendored activity library change, integrate the deliverypool changes into transportcontroller
* hook up queue deletion logic
* support deleting queued http requests by target ID
* don't index APRequest by hostname in the queue
* use gorun
* use the original context's values when wrapping msg type as delivery{}
* actually log in the AP delivery worker ...
* add uncommitted changes
* use errors.AsV2()
* use errorsv2.AsV2()
* finish adding some code comments, add bad host handling to delivery workers
* slightly tweak deliveryworkerpool API, use advanced sender multiplier
* remove PopCtx() method, let others instead rely on Wait()
* shuffle things around to move delivery stuff into transport/ subpkg
* remove dead code
* formatting
* validate request before queueing for delivery
* finish adding code comments, fix up backoff code
* finish adding more code comments
* clamp minimum no. senders to 1
* add start/stop logging to delivery worker, some slight changes
* remove double logging
* use worker ptrs
* expose the embedded log fields in httpclient.Request{}
* ensure request context values are preserved when updating ctx
* add delivery worker tests
* fix linter issues
* ensure delivery worker gets inited in testrig
* fix tests to delivering messages to check worker delivery queue
* update error type to use ptr instead of value receiver
* fix test calling Workers{}.Start() instead of testrig.StartWorkers()
* update docs for advanced-sender-multiplier
* update to the latest activity library version
* add comment about not using httptest.Server{}
* update activity library so dereferencer returns full response and checks *final* link to allow for redirects
* temporarily add bodged fixed library
* remove unused code
* update getAccountFeatured() to use dereferenceCollectionPage()
* make sure to release map
* perform a 2nd decode to ensure reader is empty after primary decode
* add comment explaining choice of using Decode() instead of Unmarshal()
* update embedded activity library to latest matching https://github.com/superseriousbusiness/activity/pull/21
* add checks to look for changed URI and re-check database if redirected
* update max iteration count to 512, add checks during dereferenceAncestors() for indirect URLs
* remove doubled-up code
* fix use of status instead of current
* use URIs for checking equality for security
* use the latest known URI for boost_of_uri in case original was an indirect
* add dereferenceCollection() function for dereferenceAccountFeatured()
* pull in latest github.com/superseriousbusiness/activity version (and remove the bodge!!)
* fix typo in code comments
* update decodeType() to accept a readcloser and handle body closing
* switch to checking using BoostOfID and add note why not using BoostOfURI
* ensure InReplyTo gets unset when deleting status parent in case currently stubbed
* add tests for Collection and CollectionPage iterators
On outgoing `GET` requests that are signed (e.g. authorized fetch),
if the initial request fails with `401`, try again, but _without_
the query parameters included in the HTTP signature.
This is primarily useful for compatibility with Mastodon; though
hopefully this can be removed in the not-too-distant future, as
they've started changing their behavior here.
Signed-off-by: Milas Bowman <devnull@milas.dev>
* [chore] Simplify the User-Agent string
RFC 9110[1] includes a definition for the format of a user-agent header:
User-Agent = product *( RWS ( product / comment ) )
product = token ["/" product-version]
product-version = token
comment = "(" *( ctext / quoted-pair / comment ) ")"
ctext = HTAB / SP / %x21-27 / %x2A-5B / %x5D-7E / obs-text
An example given in the RFC: User-Agent: CERN-LineMode/2.15 libwww/2.17b3
The idea is typically start with the most important product/version, add
a (comment) if necessary and then include any auxilliary products.
However, the RFC warns against including too many auxiliary products as
those can be unnecessarily revealing.
For automated systems (i.e not a browser), the common and recommended
format is <product></version> (+uri-for-contact), followed with any
additional <product>/<version> pairs that are relevant.
This changes our UA to match that convention more closely. This makes it
easier for administrators who do user-agent parsing for statistics or
other purposes to correctly identify the version of GoToSocial.
Currently tools tend to get confused by the lack of a /<version> on the
start of our string.
[1]: https://www.rfc-editor.org/rfc/rfc9110.html#name-user-agents
* [chore] Don't use app name in UA
From all the GtS UAs I've collected, nobody seems to set/change this,
so we might as well use the static string.
The main usefulness for this is when you have multilpe GtS instances
connecting to the same DB, so they can identify as different instances
by changing the application name (though it should already be obvious
from having different usernames).
* rewrite cache library as codeberg.org/gruf/go-structr, implement in gotosocial
* use actual go-structr release version (not just commit hash)
* revert go toolchain changes (damn you go for auto changing this)
* fix go mod woes
* ensure %w is used in calls to errs.Appendf()
* fix error checking
* fix possible panic
* remove unnecessary start/stop functions, move to main Cache{} struct, add note regarding which caches require start/stop
* fix copy-paste artifact... 😇
* fix all comment copy-paste artifacts
* remove dropID() function, now we can just use slices.DeleteFunc()
* use util.Deduplicate() instead of collate(), move collate to util
* move orderByIDs() to util package and "generify"
* add a util.DeleteIf() function, use this to delete entries on failed population
* use slices.DeleteFunc() instead of util.DeleteIf() (i had the logic mixed up in my head somehow lol)
* add note about how collate differs from deduplicate
* [bugfix] process account delete side effects in serial, not in parallel
* StartWorkers / StartNoopWorkers for tests
* undo testrig trace logging
* log errors instead of immediately returning
* update typeconverter to use state structure
* deinterface the typeutils.TypeConverter -> typeutils.Converter
* finish copying over old type converter code comments
* fix cherry-pick merge issues, fix tests pointing to old typeutils interface type still
* add automatic cache max size generation based on ratios of a singular fixed memory target
Signed-off-by: kim <grufwub@gmail.com>
* remove now-unused cache max-size config variables
Signed-off-by: kim <grufwub@gmail.com>
* slight ratio tweak
Signed-off-by: kim <grufwub@gmail.com>
* remove unused visibility config var
Signed-off-by: kim <grufwub@gmail.com>
* add secret little ratio config trick
Signed-off-by: kim <grufwub@gmail.com>
* fixed a word
Signed-off-by: kim <grufwub@gmail.com>
* update cache library to remove use of TTL in result caches + slice cache
Signed-off-by: kim <grufwub@gmail.com>
* update other cache usages to use correct interface
Signed-off-by: kim <grufwub@gmail.com>
* update example config to explain the cache memory target
Signed-off-by: kim <grufwub@gmail.com>
* update env parsing test with new config values
Signed-off-by: kim <grufwub@gmail.com>
* do some ratio twiddling
Signed-off-by: kim <grufwub@gmail.com>
* add missing header
* update envparsing with latest defaults
Signed-off-by: kim <grufwub@gmail.com>
* update size calculations to take into account result cache, simple cache and extra map overheads
Signed-off-by: kim <grufwub@gmail.com>
* tweak the ratios some more
Signed-off-by: kim <grufwub@gmail.com>
* more nan rampaging
Signed-off-by: kim <grufwub@gmail.com>
* fix envparsing script
Signed-off-by: kim <grufwub@gmail.com>
* update cache library, add sweep function to keep caches trim
Signed-off-by: kim <grufwub@gmail.com>
* sweep caches once a minute
Signed-off-by: kim <grufwub@gmail.com>
* add a regular job to sweep caches and keep under 80% utilisation
Signed-off-by: kim <grufwub@gmail.com>
* remove dead code
Signed-off-by: kim <grufwub@gmail.com>
* add new size library used to libraries section of readme
Signed-off-by: kim <grufwub@gmail.com>
* add better explanations for the mem-ratio numbers
Signed-off-by: kim <grufwub@gmail.com>
* update go-cache
Signed-off-by: kim <grufwub@gmail.com>
* library version bump
Signed-off-by: kim <grufwub@gmail.com>
* update cache.result{} size model estimation
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
For some reason we hit the case in CI where the
TestFingerWithHostMetaCacheStrategy seems to experience some time
dilation. It's possible this is a genuine bug, but I can't for the life
of me reproduce it locally, even after having run this test thousands of
times (-count=1000 when invoking go test etc.)
This changes the test to explicitly stop the webfinger cache, set TTL
and Sweep frequency to something well beyond the lifetime of the cache
during the test and then starts the cache again. Hopefully that does it,
because the other option that remains is that for some reason
timekeeping in CI/Docker is not as precise as when running the test on a
host.
* update go-cache library
Signed-off-by: kim <grufwub@gmail.com>
* fix broken test after cache library upgrade
Signed-off-by: kim <grufwub@gmail.com>
* fix the webfinger test
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
* media manager tidy-up: de-interface and remove unused PostDataFunc
Signed-off-by: kim <grufwub@gmail.com>
* remove last traces of media.Manager being an interface
Signed-off-by: kim <grufwub@gmail.com>
* update error to provide caller, allow tuneable via build tags
Signed-off-by: kim <grufwub@gmail.com>
* remove kim-specific build script changes
Signed-off-by: kim <grufwub@gmail.com>
* fix merge conflicts
Signed-off-by: kim <grufwub@gmail.com>
* update build-script to support externally setting build variables
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
* start working on lists
* further list work
* test list db functions nicely
* more work on lists
* peepoopeepoo
* poke
* start list timeline func
* we're getting there lads
* couldn't be me working on stuff... could it?
* hook up handlers
* fiddling
* weeee
* woah
* screaming, pissing
* fix streaming being a whiny baby
* lint, small test fix, swagger
* tidying up, testing
* fucked! by the linter
* move timelines to state like a boss
* add timeline start to tests using state
* invalidate lists
* add back removed ValidateRequest() before backoff-retry loop
Signed-off-by: kim <grufwub@gmail.com>
* include response body in error response log
Signed-off-by: kim <grufwub@gmail.com>
* improved error response body draining
Signed-off-by: kim <grufwub@gmail.com>
* add more code commenting
Signed-off-by: kim <grufwub@gmail.com>
* move new error response logic to gtserror, handle instead in transport.Transport{} impl
Signed-off-by: kim <grufwub@gmail.com>
* appease ye oh mighty linter
Signed-off-by: kim <grufwub@gmail.com>
* fix mockhttpclient not setting request in http response
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
Every now and then the TestFingerWithHostMetaCacheStrategy would fail on
a time related error. I suspect suite.Equal doesn't quite work as
expected when given two time.Time's, so instead explicitly check with
the time.Equal.
* [bugfix] Fix NegotiateAccept with multi accept
There's a bug in Gin's NegotiateFormat that doesn't handle the presence
of multilpe accept headers. This lifts the code from the PR @tsmethurst
sent a year ago to Gin into our codebase to fix the issue.
* [bugfix] Concat accept header in webfinger
Some implementations bug out when there's multiple accept headers,
including Gin (see 7050112af1). But things
seem to work reliably with a single accept header with multiple parts.
Fixes: #1793
* check for tls, x509 errors using string.Contains() since crypto/tls sucks
Signed-off-by: kim <grufwub@gmail.com>
* use 2* maxprocs
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
* revamp http client to not limit requests, instead use sender worker
Signed-off-by: kim <grufwub@gmail.com>
* remove separate sender worker pool, spawn 2*GOMAXPROCS batch senders each time, no need for transport cache sweeping
Signed-off-by: kim <grufwub@gmail.com>
* improve batch senders to keep popping recipients until remote URL found
Signed-off-by: kim <grufwub@gmail.com>
* fix recipient looping issue
Signed-off-by: kim <grufwub@gmail.com>
* fix missing mutex unlock
Signed-off-by: kim <grufwub@gmail.com>
* move request id ctx key to gtscontext, finish filling out more code comments, add basic support for not logging client IP
Signed-off-by: kim <grufwub@gmail.com>
* slight code reformatting
Signed-off-by: kim <grufwub@gmail.com>
* a whitespace
Signed-off-by: kim <grufwub@gmail.com>
* remove unused code
Signed-off-by: kim <grufwub@gmail.com>
* add missing license headers
Signed-off-by: kim <grufwub@gmail.com>
* fix request backoff calculation
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
* [chore] Remove years from all license headers
Years or year ranges aren't required in license headers. Many projects
have removed them in recent years and it avoids a bit of yearly toil.
In many cases our copyright claim was also a bit dodgy since we added
the 2021-2023 header to files created after 2021 but you can't claim
copyright into the past that way.
* [chore] Add license header check
This ensures a license header is always added to any new file. This
avoids maintainers/reviewers needing to remember to check for and ask
for it in case a contribution doesn't include it.
* [chore] Add missing license headers
* [chore] Further updates to license header
* Use the more common // indentend comment format
* Remove the hack we had for the linter now that we use the // format
* Add SPDX license identifier
When we receive an HTTP 410 on webfinger it means the resource we asked
for (the account) is gone, but the endpoint itself responded. In such
cases we want to treat the request as successful from a cache (renewal)
point of view, while still returning an error from Finger.
Follow-up for #1588
* [feature] Discover webfinger through host-meta
This implements a fallback for discovering the webfinger endpoint in
case the /.well-known/webfinger endpoint wasn't properly redirected.
Some instances do this because the recommendation used to be to use
host-meta for the webfinger redirect in the before times.
Closes#1558.
* [bug] Ensure we only ever update cache on success
* [chore] Move finger tests to their own place
This adds a test suite for transport and moves the finger cache tests
into there instead of abusing the search test suite.
* [chore] cleanup the test a bit more
We don't really need a separate function for the oddly located webfinger
response as we check the full URL string anyway
* Address review comments
* [chore] update config example
* [chore] access DB only through state in controller
* add error value wrapping, include status code / not found flags from transport errors, update error usages
Signed-off-by: kim <grufwub@gmail.com>
* add code commenting for gtserror functions
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
* improve error readability, mark "bad hosts" as fastFail
Signed-off-by: kim <grufwub@gmail.com>
* pull in latest go-byteutil version with byteutil.Reader{}
Signed-off-by: kim <grufwub@gmail.com>
* use rewindable body reader for post requests
Signed-off-by: kim <grufwub@gmail.com>
---------
Signed-off-by: kim <grufwub@gmail.com>
This adds a lightweight form of tracing to GTS. Each incoming request is
assigned a Request ID which we then pass on and log in all our log
lines. Any function that gets called downstream from an HTTP handler
should now emit a requestID=value pair whenever it logs something.
Co-authored-by: kim <grufwub@gmail.com>
* remove throttling timers, support setting retry-after, use retry-after in transport
* remove unused variables
* add throttling-retry-after to cmd flags
* update envparsing to include new throttling-retry-after
* update example config to include retry-after documentation
* also support retry-after formatted as date-time, ensure max backoff time
---------
Signed-off-by: kim <grufwub@gmail.com>
