stackage-server/src/Stackage/Database/Cron.hs
2025-02-27 15:05:44 +02:00

934 lines
44 KiB
Haskell

{-# LANGUAGE CPP #-}
{-# LANGUAGE LambdaCase #-}
{-# LANGUAGE NamedFieldPuns #-}
{-# LANGUAGE NoImplicitPrelude #-}
{-# LANGUAGE RecordWildCards #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE TupleSections #-}
{-# LANGUAGE FlexibleContexts #-}
module Stackage.Database.Cron
( stackageServerCron
, newHoogleLocker
, singleRun
, StackageCronOptions(..)
, defHaddockBucketName
, defHaddockBucketUrl
) where
import Conduit
import Control.DeepSeq
import Control.SingleRun
import Control.Lens ((?~))
import qualified Data.ByteString.Char8 as BS8
import qualified Data.Conduit.Binary as CB
import Data.Conduit.Tar (FileInfo(..), FileType(..), untar)
import Data.Conduit.Zlib (WindowBits(WindowBits), compress, ungzip)
import qualified Data.IntMap.Strict as IntMap
import Data.Monoid (Any(..))
import Data.Streaming.Network (bindPortTCP)
import Data.Yaml (decodeFileEither)
import Database.Persist hiding (exists)
import Database.Persist.Postgresql hiding (exists)
import qualified Hoogle
import Amazonka hiding (Request, length, error)
import Amazonka.Data.Text (toText)
import Amazonka.S3
import Amazonka.S3.ListObjectsV2
import Amazonka.S3.Lens
import Amazonka.Lens
import Network.HTTP.Client
import Network.HTTP.Client.Conduit (bodyReaderSource)
import Network.HTTP.Simple (getResponseBody, httpJSONEither)
import Network.HTTP.Types (status200, status404)
import Pantry (CabalFileInfo(..), DidUpdateOccur(..),
HpackExecutable(HpackBundled), PackageIdentifierRevision(..),
defaultCasaMaxPerRequest, defaultCasaRepoPrefix,
defaultHackageSecurityConfig, defaultSnapshotLocation)
import Pantry.Internal.Stackage (HackageTarballResult(..), PantryConfig(..),
Storage(..), forceUpdateHackageIndex,
getHackageTarball, packageTreeKey)
import Path (parseAbsDir, toFilePath)
import RIO
import RIO.Directory
import RIO.File
import RIO.FilePath
import RIO.List as L
import qualified RIO.Map as Map
import RIO.Process (mkDefaultProcessContext)
import qualified RIO.Set as Set
import qualified RIO.Text as T
import RIO.Time
import Settings
import Stackage.Database.Github
import Stackage.Database.PackageInfo
import Stackage.Database.Query
import Stackage.Database.Schema
import Stackage.Database.Types
import System.Environment (getEnvironment)
import UnliftIO.Concurrent (getNumCapabilities)
import Web.PathPieces (fromPathPiece, toPathPiece)
import qualified Control.Retry as Retry
hoogleKey :: SnapName -> Text
hoogleKey name = T.concat
[ "hoogle/"
, toPathPiece name
, "/"
, VERSION_hoogle
, ".hoo"
]
hoogleUrl :: SnapName -> Text -> Text
hoogleUrl n haddockBucketUrl = T.concat
[ haddockBucketUrl
, "/"
, hoogleKey n
]
hackageDeprecatedUrl :: Request
hackageDeprecatedUrl = "https://hackage.haskell.org/packages/deprecated.json"
withStorage :: (Storage -> IO a) -> IO a
withStorage inner = do
as <- getAppSettings
withStackageDatabase False (appDatabase as) (\db -> inner (Storage (runDatabase db) id))
getStackageSnapshotsDir :: RIO StackageCron FilePath
getStackageSnapshotsDir = do
cron <- ask
cloneOrUpdate (scStackageRoot cron) (scSnapshotsRepo cron)
withResponseUnliftIO :: MonadUnliftIO m => Request -> Manager -> (Response BodyReader -> m b) -> m b
withResponseUnliftIO req man f = withRunInIO $ \ runInIO -> withResponse req man (runInIO . f)
-- | Returns an action that, under the SingleRun wrapper that ensures only one
-- thing at a time is writing the file in question, ensure that a Hoogle
-- database exists on the filesystem for the given SnapName. But only going so
-- far as downloading it from the haddock bucket. See 'buildAndUploadHoogleDBs' for the
-- function that puts it there in the first place. If no db exists in the
-- bucket, the action will return 'Nothing'.
--
-- The location searched is $PWD/hoogle/<snapshot>/<hoogle-version>.hoo
-- E.g. in production, ~stackage-update/hoogle/lts-22.20/5.0.18.4.hoo (for stackage-update).
newHoogleLocker ::
(HasLogFunc env, MonadIO m) => env -> Manager -> Text -> m (SingleRun SnapName (Maybe FilePath))
newHoogleLocker env man bucketUrl = mkSingleRun hoogleLocker
where
hoogleLocker :: MonadIO n => SnapName -> n (Maybe FilePath)
hoogleLocker name =
runRIO env $ do
let fp = T.unpack $ hoogleKey name
exists <- doesFileExist fp
if exists
then return $ Just fp
else do
req' <- parseRequest $ T.unpack $ hoogleUrl name bucketUrl
let req = req' {decompress = const False}
withResponseUnliftIO req man $ \res ->
case responseStatus res of
status
| status == status200 -> do
createDirectoryIfMissing True $ takeDirectory fp
withBinaryFileDurableAtomic fp WriteMode $ \h ->
runConduitRes $
bodyReaderSource (responseBody res) .| ungzip .|
sinkHandle h
return $ Just fp
| status == status404 -> do
logWarn $ "NotFound: " <> display (hoogleUrl name bucketUrl)
return Nothing
| otherwise -> do
body <- liftIO $ brConsume $ responseBody res
logWarn $ "Unexpected status: " <> displayShow status
mapM_ (logWarn . displayBytesUtf8) body
return Nothing
getHackageDeprecations ::
(HasLogFunc env, MonadReader env m, MonadIO m) => m [Deprecation]
getHackageDeprecations = do
let policy = Retry.exponentialBackoff 50 <> Retry.limitRetries 5
jsonResponseDeprecated <-
liftIO $ Retry.recoverAll policy $ const $ httpJSONEither hackageDeprecatedUrl
case getResponseBody jsonResponseDeprecated of
Left err -> do
logError $
"There was an error parsing deprecated.json file: " <>
fromString (displayException err)
return []
Right deprecated -> return deprecated
stackageServerCron :: StackageCronOptions -> IO ()
stackageServerCron StackageCronOptions {..} = do
void $
-- Hacky approach instead of PID files
catchIO (bindPortTCP 17834 "127.0.0.1") $
const $ throwString "Stackage Cron loader process already running, exiting."
connectionCount <- getNumCapabilities
withStorage $ \storage -> do
lo <- logOptionsHandle stdout True
stackageRootDir <- getAppUserDataDirectory "stackage"
pantryRootDir <- parseAbsDir (stackageRootDir </> "pantry")
createDirectoryIfMissing True (toFilePath pantryRootDir)
updateRef <- newMVar True
cabalImmutable <- newIORef Map.empty
cabalMutable <- newIORef Map.empty
gpdCache <- newIORef IntMap.empty
defaultProcessContext <- mkDefaultProcessContext
aws <- do
aws' <- newEnv discover
endpoint <- lookup "AWS_S3_ENDPOINT" <$> getEnvironment
pure $ case endpoint of
Nothing -> aws'
Just ep -> configureService (setEndpoint True (BS8.pack ep) 443 Amazonka.S3.defaultService) aws'
withLogFunc (setLogMinLevel scoLogLevel lo) $ \logFunc -> do
let pantryConfig =
PantryConfig
{ pcHackageSecurity = defaultHackageSecurityConfig
, pcHpackExecutable = HpackBundled
, pcRootDir = pantryRootDir
, pcStorage = storage
, pcUpdateRef = updateRef
, pcParsedCabalFilesRawImmutable = cabalImmutable
, pcParsedCabalFilesMutable = cabalMutable
, pcConnectionCount = connectionCount
, pcCasaRepoPrefix = defaultCasaRepoPrefix
, pcCasaMaxPerRequest = defaultCasaMaxPerRequest
, pcSnapshotLocation = defaultSnapshotLocation
}
currentHoogleVersionId <- runRIO logFunc $ do
runStackageMigrations' pantryConfig
getCurrentHoogleVersionIdWithPantryConfig pantryConfig
let stackage =
StackageCron
{ scPantryConfig = pantryConfig
, scStackageRoot = stackageRootDir
, scProcessContext = defaultProcessContext
, scLogFunc = logFunc
, scForceFullUpdate = scoForceUpdate
, scCachedGPD = gpdCache
, scEnvAWS = aws
, scDownloadBucketName = scoDownloadBucketName
, scDownloadBucketUrl = scoDownloadBucketUrl
, scUploadBucketName = scoUploadBucketName
, scSnapshotsRepo = scoSnapshotsRepo
, scReportProgress = scoReportProgress
, scCacheCabalFiles = scoCacheCabalFiles
, scHoogleVersionId = currentHoogleVersionId
}
runRIO stackage (runStackageUpdate scoDoNotUpload)
runStackageUpdate :: Bool -> RIO StackageCron ()
runStackageUpdate doNotUpload = do
forceFullUpdate <- scForceFullUpdate <$> ask
logInfo $ "Starting stackage-cron update" <> bool "" " with --force-update" forceFullUpdate
runStackageMigrations
didUpdate <- forceUpdateHackageIndex (Just "stackage-server cron job")
case didUpdate of
UpdateOccurred -> logInfo "Updated hackage index"
NoUpdateOccurred -> logInfo "No new packages in hackage index"
logInfo "Getting deprecated info now"
getHackageDeprecations >>= setDeprecations
corePackageGetters <- makeCorePackageGetters
runResourceT $
join $
-- @createOrUpdateSnapshot@ processes package N while processing docs for
-- package N-1. This @pure ()@ is the "documentation processing action"
-- for the -1'th package.
runConduit $ sourceSnapshots .| foldMC (createOrUpdateSnapshot corePackageGetters) (pure ())
unless doNotUpload uploadSnapshotsJSON
buildAndUploadHoogleDBs doNotUpload
logInfo "Finished building and uploading Hoogle DBs"
-- | This will look at 'global-hints.yaml' and will create core package getters that are reused
-- later for adding those package to individual snapshots.
makeCorePackageGetters ::
RIO StackageCron (Map CompilerP [CorePackageGetter])
makeCorePackageGetters = do
rootDir <- scStackageRoot <$> ask
contentDir <- getStackageContentDir rootDir
backupCoreCabalFiles <- getBackupCoreCabalFiles rootDir
liftIO (decodeFileEither (contentDir </> "stack" </> "global-hints.yaml")) >>= \case
Right (hints :: Map CompilerP (Map PackageNameP VersionP)) ->
Map.traverseWithKey
(\compiler ->
fmap Map.elems .
Map.traverseMaybeWithKey (makeCorePackageGetter compiler backupCoreCabalFiles))
hints
Left exc -> do
logError $
"Error parsing 'global-hints.yaml' file: " <> fromString (displayException exc)
return mempty
-- | Packages distributed with GHC aren't taken from Hackage like normal
-- packages. Release managers do upload them, however, so that their docs are
-- available.
--
-- Or at least, they should. The release process was fragile, and some packages
-- weren't uploaded. This mechanism gives us a chance to fill in missing
-- packages.
getBackupCoreCabalFiles ::
FilePath
-> RIO StackageCron (Map PackageIdentifierP (GenericPackageDescription, CabalFileIds))
getBackupCoreCabalFiles rootDir = do
backupCoreCabalFilesDir <- getBackupCoreCabalFilesDir rootDir
cabalFileNames <- getDirectoryContents backupCoreCabalFilesDir
cabalFiles <-
forM (filter (isExtensionOf ".cabal") cabalFileNames) $ \cabalFileName ->
let pidTxt = T.pack (dropExtension (takeFileName cabalFileName))
in case fromPathPiece pidTxt of
Nothing -> do
logError $ "Invalid package identifier: " <> fromString cabalFileName
pure Nothing
Just pid -> do
cabalBlob <- readFileBinary (backupCoreCabalFilesDir </> cabalFileName)
mCabalInfo <- run $ addCabalFile pid cabalBlob
pure ((,) pid <$> mCabalInfo)
pure $ Map.fromList $ catMaybes cabalFiles
-- | Core package info rarely changes between the snapshots, therefore it would be wasteful to
-- load, parse and update all packages from gloabl-hints for each snapshot. Instead we produce
-- a memoized version that will do it once initially and then return information about a
-- package on subsequent invocations.
--
-- FIXME: The compiler argument is unused (and never has been). Should it be used?
makeCorePackageGetter ::
CompilerP
-> Map PackageIdentifierP (GenericPackageDescription, CabalFileIds)
-> PackageNameP
-> VersionP
-> RIO StackageCron (Maybe CorePackageGetter)
makeCorePackageGetter _compiler fallbackCabalFileMap pname ver =
run (getHackageCabalByRev0 pid) >>= \case
Nothing -> do
logWarn $
"Core package from global-hints: '" <> display pid <> "' was not found in pantry."
forM (Map.lookup pid fallbackCabalFileMap) $ \(gpd, cabalFileIds) -> do
logInfo $
"Falling back on '" <> display pid <>
".cabal' file from the commercialhaskell/core-cabal-files repo"
pure $ pure (Left cabalFileIds, Nothing, pid, gpd)
Just (hackageCabalId, blobId, _) -> do
pkgInfoRef <- newIORef Nothing -- use for caching of pkgInfo
let getCabalFileIdsTree gpd =
\case
Just tree -> pure $ Right tree
Nothing -> Left <$> getCabalFileIds blobId gpd
let getMemoPackageInfo =
readIORef pkgInfoRef >>= \case
Just pkgInfo -> return pkgInfo
Nothing -> do
whenM (scReportProgress <$> ask) $
logSticky $ "Loading core package: " <> display pid
-- I have no idea what's happening here. I guess I
-- don't know what it means to "load" a package.
-- What is actually going on?
htr <- getHackageTarball pir Nothing
case htrFreshPackageInfo htr of
Just (gpd, treeId) -> do
eTree <-
run $ do
mTree <- getEntity treeId
getCabalFileIdsTree gpd mTree
let pkgInfo = (eTree, Just hackageCabalId, pid, gpd)
gpd `deepseq` writeIORef pkgInfoRef $ Just pkgInfo
pure pkgInfo
Nothing -> do
(gpd, eCabalTree) <-
run $ do
cabalBlob <- loadBlobById blobId
let gpd = parseCabalBlob cabalBlob
mTree <- getTreeForKey (packageTreeKey (htrPackage htr))
(,) gpd <$> getCabalFileIdsTree gpd mTree
let pkgInfo = (eCabalTree, Just hackageCabalId, pid, gpd)
gpd `deepseq` writeIORef pkgInfoRef $ Just pkgInfo
pure pkgInfo
pure $ Just getMemoPackageInfo
where
pid = PackageIdentifierP pname ver
pir =
PackageIdentifierRevision (unPackageNameP pname) (unVersionP ver) (CFIRevision (Revision 0))
-- | Populates the database with information about a package?
--
-- Specifically, a pantry package is being added to a particular snapshot.
--
-- Extra information like compiler and flags are passed on in order to properly
-- figure out dependencies and modules.
--
-- TODO: for now it is only from hackage. PantryPackage needs an update to use other origins
addPantryPackage ::
SnapshotId -> CompilerP -> Bool -> Map FlagNameP Bool -> PantryPackage -> RIO StackageCron Bool
addPantryPackage snapId compiler isHidden flags (PantryPackage pcabal pTreeKey) = do
env <- ask
let pkgDescCache = scCachedGPD env
cacheP = scCacheCabalFiles env
let blobKeyToInt = fromIntegral . unSqlBackendKey . unBlobKey
let cachedPkgDesc cabalBlobId pkgDesc =
pkgDesc `deepseq`
atomicModifyIORef' pkgDescCache (\cacheMap -> (IntMap.insert cabalBlobId pkgDesc cacheMap, pkgDesc))
let getPkgDesc cabalBlobId =
\case
Just pkgDesc | cacheP -> cachedPkgDesc (blobKeyToInt cabalBlobId) pkgDesc
Just pkgDesc -> pure pkgDesc
Nothing | cacheP -> do
cacheMap <- readIORef pkgDescCache
case IntMap.lookup (blobKeyToInt cabalBlobId) cacheMap of
Just pkgDesc -> pure pkgDesc
Nothing ->
loadBlobById cabalBlobId >>=
cachedPkgDesc (blobKeyToInt cabalBlobId) . parseCabalBlob
Nothing -> parseCabalBlob <$> loadBlobById cabalBlobId
let storeHackageSnapshotPackage hackageCabalId mTreeId mpkgDesc =
getTreeForKey pTreeKey >>= \case
-- error case #1
Just (Entity treeId' _)
| Just treeId <- mTreeId
, treeId /= treeId' -> do
lift $ logError $ "Pantry Tree Key mismatch for: " <> display pcabal
pure False
-- happy case
Just pkgTree@(Entity _ Tree {treeCabal})
| Just cabalBlobId <- treeCabal -> do
pkgDesc <- getPkgDesc cabalBlobId mpkgDesc
addSnapshotPackage snapId compiler Hackage (Right pkgTree) (Just hackageCabalId) isHidden flags packageId pkgDesc
pure True
-- error case #2
_ -> do
lift $ logError $ "Pantry is missing the source tree for " <> display pcabal
pure False
mHackageCabalInfo <- run $ getHackageCabalByKey packageId (pcCabalKey pcabal)
case mHackageCabalInfo of
Nothing -> do
logError $ "Could not find the cabal file for: " <> display pcabal
pure False
Just (hackageCabalId, Nothing) -> do
mHPI <-
htrFreshPackageInfo <$>
getHackageTarball (toPackageIdentifierRevision pcabal) (Just pTreeKey)
run $
case mHPI of
Just (pkgDesc, treeId) -> storeHackageSnapshotPackage hackageCabalId (Just treeId) (Just pkgDesc)
Nothing -> storeHackageSnapshotPackage hackageCabalId Nothing Nothing
Just (hackageCabalId, mTreeId) -> run $ storeHackageSnapshotPackage hackageCabalId mTreeId Nothing
where
packageId = PackageIdentifierP (pcPackageName pcabal) (pcVersion pcabal)
-- | Download a list of available .html files from S3 bucket for a particular resolver and record
-- in the database which modules have documentation available for them.
checkForDocs :: SnapshotId -> SnapName -> ResourceT (RIO StackageCron) ()
checkForDocs snapshotId snapName = do
bucketName <- lift (scDownloadBucketName <$> ask)
env <- asks scEnvAWS
-- it is faster to download all modules in this snapshot separately, rather
-- than process them with a conduit all the way to the database.
packageModules <-
runConduit $
paginate env (listSnapshotObjects bucketName)
.| concatMapC (fromMaybe [] . (^. listObjectsV2Response_contents))
.| mapC (\obj -> toText (obj ^. object_key))
.| concatMapC (T.stripSuffix ".html" >=> T.stripPrefix prefix >=> pathToPackageModule)
.| sinkList
-- Cache SnapshotPackageId rather than look it up many times for each module in the package.
sidsCacheRef <- newIORef Map.empty
-- The other half of the cores are used in 'updateSnapshot'
n <- max 1 . (`div` 2) <$> getNumCapabilities
unexpectedPackages <- lift $ pooledMapConcurrentlyN n (markModule sidsCacheRef) packageModules
forM_ (Set.fromList $ catMaybes unexpectedPackages) $ \pid ->
lift $ logWarn $
"Documentation found for package '" <> display pid <>
"', which does not exist in this snapshot: " <>
display snapName
where
prefix = textDisplay snapName <> "/"
listSnapshotObjects bucketName = newListObjectsV2 (BucketName bucketName) & listObjectsV2_prefix ?~ prefix
-- | This function records all package modules that have documentation available, the ones
-- that are not found in the snapshot reported back as an error. Besides being run
-- concurrently this function optimizes the SnapshotPackageId lookup as well, since that can
-- be shared amongst many modules of one package.
markModule sidsCacheRef (pid, modName) = do
sidsCache <- readIORef sidsCacheRef
let mSnapshotPackageId = Map.lookup pid sidsCache
mFound <- run $ markModuleHasDocs snapshotId pid mSnapshotPackageId modName
case mFound of
Nothing -> pure $ Just pid -- This package doesn't exist in the snapshot!
Just snapshotPackageId
| Nothing <- mSnapshotPackageId -> do
atomicModifyIORef'
sidsCacheRef
(\cacheMap -> (Map.insert pid snapshotPackageId cacheMap, ()))
pure Nothing
_ -> pure Nothing
data SnapshotFileInfo = SnapshotFileInfo
{ sfiSnapName :: !SnapName
, sfiUpdatedOn :: !UTCTime
, sfiSnapshotFileGetter :: !(RIO StackageCron (Maybe SnapshotFile))
}
-- | Use 'github.com/commercialhaskell/stackage-snapshots' repository to source all of the packages
-- one snapshot at a time.
sourceSnapshots :: ConduitT a SnapshotFileInfo (ResourceT (RIO StackageCron)) ()
sourceSnapshots = do
snapshotsDir <- lift $ lift getStackageSnapshotsDir
sourceDirectoryDeep False (snapshotsDir </> "lts") .| concatMapMC (getLtsParser snapshotsDir)
sourceDirectoryDeep False (snapshotsDir </> "nightly") .| concatMapMC (getNightlyParser snapshotsDir)
where
makeSnapshotFileInfo gitDir fp mFileNameDate snapName = do
let parseSnapshot updatedOn = do
esnap <- liftIO $ decodeFileEither fp
case esnap of
Right snap ->
let publishDate =
sfPublishDate snap <|> mFileNameDate <|> Just (utctDay updatedOn)
in return $ Just snap {sfPublishDate = publishDate}
Left exc -> do
logError $
"Error parsing snapshot file: " <> fromString fp <> "\n" <>
fromString (displayException exc)
return Nothing
mUpdatedOn <- lastGitFileUpdate gitDir fp
forM mUpdatedOn $ \updatedOn -> do
env <- lift ask
return $
SnapshotFileInfo
{ sfiSnapName = snapName
, sfiUpdatedOn = updatedOn
, sfiSnapshotFileGetter = runRIO env (parseSnapshot updatedOn)
}
getLtsParser gitDir fp =
case mapM (BS8.readInt . BS8.pack) $ take 2 $ reverse (splitPath fp) of
Just [(minor, ".yaml"), (major, "/")] ->
makeSnapshotFileInfo gitDir fp Nothing $ SNLts major minor
_ -> do
logError
("Couldn't parse the filepath into an LTS version: " <> display (T.pack fp))
return Nothing
getNightlyParser gitDir fp =
case mapM (BS8.readInt . BS8.pack) $ take 3 $ reverse (splitPath fp) of
Just [(day, ".yaml"), (month, "/"), (year, "/")]
| Just date <- fromGregorianValid (fromIntegral year) month day ->
makeSnapshotFileInfo gitDir fp (Just date) $ SNNightly date
_ -> do
logError
("Couldn't parse the filepath into a Nightly date: " <> display (T.pack fp))
return Nothing
data DecisionResult a e = NothingToDo | NoSnapshotFile | NeedsUpdate a e | DoesntExist e
-- | Creates a new `Snapshot` if it is not yet present in the database, and decides if update
-- is necessary when it already exists.
--
-- TODO: Silently ignoring snapshots where the getter returns Nothing seems like
-- a potential problem. Anyway I'd rather run it beforehand!
decideOnSnapshotUpdate :: SnapshotFileInfo -> RIO StackageCron (Maybe (SnapshotId, SnapshotFile))
decideOnSnapshotUpdate SnapshotFileInfo {sfiSnapName, sfiUpdatedOn, sfiSnapshotFileGetter} = do
forceUpdate <- scForceFullUpdate <$> ask
let mkLogMsg rest = "Snapshot with name: " <> display sfiSnapName <> " " <> rest
mKeySnapFile <-
run (getBy (UniqueSnapshot sfiSnapName)) >>= \case
-- exists, up to date, no force-updated requested; nothing to do
Just (Entity _key snap)
| snapshotUpdatedOn snap == Just sfiUpdatedOn && not forceUpdate ->
return NothingToDo
-- exists but updatedOn was not previously set.
Just entity@(Entity _key snap)
| Nothing <- snapshotUpdatedOn snap -> do
logWarn $ mkLogMsg "did not finish updating last time."
maybe NoSnapshotFile (NeedsUpdate entity) <$> sfiSnapshotFileGetter
-- exists, but updatedOn does not match or force-update was requested.
Just entity -> do
unless forceUpdate $ logWarn $ mkLogMsg "was updated, applying new patch."
maybe NoSnapshotFile (NeedsUpdate entity) <$> sfiSnapshotFileGetter
-- does not exist
Nothing -> maybe NoSnapshotFile DoesntExist <$> sfiSnapshotFileGetter
-- Add new snapshot to the database, when necessary
case mKeySnapFile of
NothingToDo -> Nothing <$ logInfo (mkLogMsg "already exists and is up to date.")
NoSnapshotFile -> Nothing <$ logWarn (mkLogMsg "has no (readable?) snapshot file.")
NeedsUpdate (Entity oldSnapKey oldSnap) sf@SnapshotFile {sfCompiler, sfPublishDate}
| Just publishDate <- sfPublishDate -> do
let updatedSnap =
Snapshot sfiSnapName sfCompiler publishDate (snapshotUpdatedOn oldSnap)
run $ replace oldSnapKey updatedSnap
pure $ Just (oldSnapKey, sf)
| otherwise -> return Nothing
DoesntExist sf@SnapshotFile {sfCompiler, sfPublishDate}
| Just publishDate <- sfPublishDate -> do
logInfo $ mkLogMsg "is new, adding to the database."
fmap (, sf) <$>
run (insertUnique (Snapshot sfiSnapName sfCompiler publishDate Nothing))
| otherwise -> Nothing <$ logWarn (mkLogMsg "has no publish date, skipping.")
type CorePackageGetter
= RIO StackageCron ( Either CabalFileIds (Entity Tree)
, Maybe HackageCabalId
, PackageIdentifierP
, GenericPackageDescription)
-- | This is an optimized version of snapshoat loading which can load a snapshot and documentation
-- info for previous snapshot at the same time. It will execute concurrently the loading of
-- current snapshot as well as an action that was passed as an argument. At the end it will return
-- an action that should be invoked in order to mark modules that have documentation available,
-- which in turn can be passed as an argument to the next snapshot loader.
-- Something something ouroboros.
--
-- Question: When do the docs for the last snapshot get loaded?
--
-- Well, this binding is called as @join $ runConduit $ foldMC (createOrUpdateSnapshot corePackageInfoGetters) (pure ())@
--
-- So the answer: the doc-loading action for the last snapshot gets returned by @runConduit $ foldMC ...@,
-- which means it gets executed by @join $ runConduit $ foldMC ...@.
--
-- Evidence:
--
-- Since @foldMC :: (a -> b -> m a) -> a -> ConduitT b o m a@, we see
--
-- @@
-- a ~ ResourceT (RIO Stackage Cron) () -- this is the doc-loading action
-- b ~ SnapshotFileInfo
-- m ~ ResourceT (RIO StackageCron)
-- @@
-- and the foldMC creates a @ConduitT SnapshotFileInfo o (ResourceT (RIO StackageCron)) (ResourceT (RIO StackageCron) ())@
--
-- TODO: It might be more efficient to just put all the actions (snapshot
-- creation and documentation writing both) on a queue and let a bunch of
-- workers chew on it. The current impl creates arbitrary synchronization points
-- with 'runConcurrently'. Granted, I don't know what a good chunk size would
-- actually be.
createOrUpdateSnapshot ::
Map CompilerP [CorePackageGetter]
-> ResourceT (RIO StackageCron) ()
-> SnapshotFileInfo
-> ResourceT (RIO StackageCron) (ResourceT (RIO StackageCron) ())
createOrUpdateSnapshot corePackageInfoGetters prevAction sfi@SnapshotFileInfo { sfiSnapName , sfiUpdatedOn } = do
finishedDocs <- newIORef False
runConcurrently
(Concurrently (prevAction >> writeIORef finishedDocs True) *>
Concurrently (lift (loadCurrentSnapshot finishedDocs)))
where
loadCurrentSnapshot finishedDocs = do
loadDocs <-
decideOnSnapshotUpdate sfi >>= \case
-- Nothing to do, and thus no docs to process
Nothing -> return $ pure ()
Just (snapshotId, snapshotFile) ->
updateSnapshot
corePackageInfoGetters
snapshotId
sfiSnapName
sfiUpdatedOn
snapshotFile
report <- scReportProgress <$> ask
when report $
unlessM (readIORef finishedDocs) $
logSticky "Still loading the docs for previous snapshot ..."
pure loadDocs
-- | Creates Lts or Nightly entity [Question(bryan): Why not do this when
-- creating the snapshot? Why is this a separate table anyway?] and updates all
-- packages in the snapshot. If any packages are missing they will be created.
-- Returns an action that will (a) check for available documentation for the
-- packages' modules and (b) mark the packages as documented when haddock is
-- present on AWS S3.
--
-- (Only after documentation has been checked will this snapshot be marked as
-- completely updated. This is required in case something goes wrong and process
-- is interrupted.)
updateSnapshot ::
Map CompilerP [CorePackageGetter]
-> SnapshotId
-> SnapName
-> UTCTime
-> SnapshotFile
-> RIO StackageCron (ResourceT (RIO StackageCron) ()) -- ^ Returns the action for processing docs
updateSnapshot corePackageGetters snapshotId snapName updatedOn SnapshotFile {..} = do
insertSnapshotName snapshotId snapName
loadedPackageCountRef <- newIORef (0 :: Int)
let totalPackages = length sfPackages
-- A wrapper for 'addPantryPackage' that extracts the package info from
-- snapshot info, increments the count of loaded packages, and reports success
-- as a Bool.
addPantryPackageWithReport pp = do
let PantryCabal {pcPackageName} = ppPantryCabal pp
isHidden = fromMaybe False (Map.lookup pcPackageName sfHidden)
flags = fromMaybe Map.empty $ Map.lookup pcPackageName sfFlags
curSucc <- addPantryPackage snapshotId sfCompiler isHidden flags pp
atomicModifyIORef' loadedPackageCountRef (\c -> (c + 1, ()))
pure curSucc
-- Leave some cores and db connections for the doc loader
n <- max 1 . (`div` 2) <$> getNumCapabilities
before <- getCurrentTime
report <- scReportProgress <$> ask
pantryUpdatesSucceeded <-
runConcurrently
(Concurrently
(when report (runProgressReporter loadedPackageCountRef totalPackages snapName)) *>
Concurrently (pooledMapConcurrentlyN n addPantryPackageWithReport sfPackages))
after <- getCurrentTime
let timeTotal = round (diffUTCTime after before)
(mins, secs) = timeTotal `quotRem` (60 :: Int)
packagePerSecond = fromIntegral ((totalPackages * 100) `div` timeTotal) / 100 :: Float
allPantryUpdatesSucceeded = and pantryUpdatesSucceeded
logInfo $
mconcat
[ "Loading snapshot '"
, display snapName
, "' was done (in "
, displayShow mins
, "min "
, displayShow secs
, "sec). With average "
, displayShow packagePerSecond
, " packages/sec. There are still docs."
]
case Map.lookup sfCompiler corePackageGetters of
Nothing -> logError $ "Hints are not found for the compiler: " <> display sfCompiler
Just _
| not allPantryUpdatesSucceeded ->
logWarn $
mconcat
[ "There was an issue loading a snapshot '"
, display snapName
, "', deferring addition of packages "
, "from global-hints until next time."
]
Just compilerCorePackages ->
forM_ compilerCorePackages $ \getCorePackageInfo -> do
(eTree, mhcid, pid, gpd) <- getCorePackageInfo
run $ addSnapshotPackage snapshotId sfCompiler Core eTree mhcid False mempty pid gpd
return $ do
checkForDocsSucceeded <-
tryAny (checkForDocs snapshotId snapName) >>= \case
Left exc -> do
logError $ "Received exception while getting the docs: " <> displayShow exc
return False
Right () -> return True
if allPantryUpdatesSucceeded &&
checkForDocsSucceeded && Map.member sfCompiler corePackageGetters
then do
lift $ snapshotMarkUpdated snapshotId updatedOn
logInfo $ "Created or updated snapshot '" <> display snapName <> "' successfully"
else logError $ "There were errors while adding snapshot '" <> display snapName <> "'"
-- | Report how many packages has been loaded so far.
runProgressReporter :: IORef Int -> Int -> SnapName -> RIO StackageCron ()
runProgressReporter loadedPackageCountRef totalPackages snapName = do
let reportProgress = do
loadedPackageCount <- readIORef loadedPackageCountRef
when (loadedPackageCount < totalPackages) $ do
logSticky $
mconcat
[ "Loading snapshot '"
, display snapName
, "' ("
, displayShow loadedPackageCount
, "/"
, displayShow totalPackages
, ")"
]
threadDelay 1000000
reportProgress
reportProgress
-- | Uploads a json file to S3 with all latest snapshots per major lts version and one nightly.
uploadSnapshotsJSON :: RIO StackageCron ()
uploadSnapshotsJSON = do
snapshots <- snapshotsJSON
uploadBucket <- scUploadBucketName <$> ask
let key = ObjectKey "snapshots.json"
uploadFromRIO key $
set putObject_acl (Just ObjectCannedACL_Public_read) $
set putObject_contentType (Just "application/json") $
newPutObject (BucketName uploadBucket) key (toBody snapshots)
-- | Writes a gzipped version of hoogle db into temporary file onto the file system and then uploads
-- it to S3. Temporary file is removed upon completion
uploadHoogleDB :: FilePath -> ObjectKey -> RIO StackageCron ()
uploadHoogleDB fp key =
withTempFile (takeDirectory fp) (takeFileName fp <.> "gz") $ \fpgz h -> do
runConduitRes $ sourceFile fp .| compress 9 (WindowBits 31) .| CB.sinkHandle h
hClose h
body <- toBody <$> readFileBinary fpgz
uploadBucket <- scUploadBucketName <$> ask
uploadFromRIO key $
-- FIXME: I should also set content encoding explicitly here. But
-- then I would break stackage-server, which applies an 'ungzip' in
-- 'newHoogleLocker'. :(
set putObject_acl (Just ObjectCannedACL_Public_read) $ newPutObject (BucketName uploadBucket) key body
uploadFromRIO :: (AWSRequest a, Typeable a, Typeable (AWSResponse a)) => ObjectKey -> a -> RIO StackageCron ()
uploadFromRIO key po = do
logInfo $ "Uploading " <> displayShow key <> " to S3 bucket."
env <- asks scEnvAWS
eres <- runResourceT $ trying _Error $ send env po
case eres of
Left e ->
logError $ "Couldn't upload " <> displayShow key <> " to S3 becuase " <> displayShow e
Right _ -> logInfo $ "Successfully uploaded " <> displayShow key <> " to S3"
-- | As the name says, build and upload Hoogle DBs.
--
-- Which DBs? The last 5 LTS and the last 5 Nightlies that are missing their
-- Hoogle DBs.
--
-- How? It downloads the Hoogle inputs that were previously generated alongside
-- the Haddocks, runs @hoogle@ on them, and uploads the result back to the same
-- bucket. Those inputs were generated by snapshot curation.
--
-- Why? I feel like this should be a short Bash script using curl and hoogle, and
-- maybe one day it will be.
--
-- This action is only run by stackage-server-cron.
buildAndUploadHoogleDBs :: Bool -> RIO StackageCron ()
buildAndUploadHoogleDBs doNotUpload = do
snapshots <- lastLtsNightlyWithoutHoogleDb 5 5
env <- ask
awsEnv <- asks scEnvAWS
bucketUrl <- asks scDownloadBucketUrl
-- locker is an action that returns the path to a hoogle db, if one exists
-- in the haddock bucket already. It takes the SnapName as an argument.
-- I think it might be overkill.
locker <- newHoogleLocker (env ^. logFuncL) (awsEnv ^. env_manager) bucketUrl
let -- These bindings undo a questionable conflation of operations
insertH = checkInsertSnapshotHoogleDb True
checkH = checkInsertSnapshotHoogleDb False
for_ snapshots $ \(snapshotId, snapName) ->
-- Even though we just got a list of snapshots that don't have hoogle
-- databases, we check again. For some reason. I don't see how this can
-- actually be useful. both lastLtsNightlyWithoutHoogleDb and
-- checkInsertSnapshotHoogleDb just check against SnapshotHoogleDb.
-- Perhaps the check can be removed.
unlessM (checkH snapshotId) $ do
logInfo $ "Starting Hoogle database download: " <> display (hoogleKey snapName)
-- Check if the database already exists (by downloading it).
-- FIXME: Why not just send a HEAD?
-- Perhaps the idea was to put the hoogle database somewhere the
-- main Stackage server process can find it? But nowadays
-- stackage-server downloads its own version separately.
mfp <- singleRun locker snapName
case mfp of
Just _ -> do
-- Something bad must have happened: we created the hoogle db
-- previously, but didn't get to record it as available.
logWarn $ "Unregistered hoogle database found for: " <> display snapName
<> ". Registering now."
void $ insertH snapshotId
Nothing -> do
logInfo $ "Current hoogle database does not yet exist in the bucket for: " <> display snapName
mfp' <- createHoogleDB snapshotId snapName
forM_ mfp' $ \fp -> do
let key = hoogleKey snapName
dest = T.unpack key
createDirectoryIfMissing True $ takeDirectory dest
renamePath fp dest
unless doNotUpload $ do
uploadHoogleDB dest (ObjectKey key)
void $ insertH snapshotId
-- | Create a hoogle db from haddocks for the given snapshot.
--
-- Haddocks are downloaded from the documentation bucket, where they were
-- uploaded as a tar file.
--
-- Returns the path to the .hoo database, which will be found in the first
-- argument. It will look like @<rootDir>/hoogle-gen/output.hoo@.
createHoogleDB :: SnapshotId -> SnapName -> RIO StackageCron FilePath
createHoogleDB snapshotId snapName = do
logInfo $ "Creating Hoogle DB for " <> display snapName
downloadBucketUrl <- scDownloadBucketUrl <$> ask
let root = "hoogle-gen"
outname = root </> "output.hoo"
tarKey = toPathPiece snapName <> "/hoogle/orig.tar"
tarUrl = downloadBucketUrl <> "/" <> tarKey
tarFP = root </> T.unpack tarKey
-- When tarball is downloaded it is saved with durability and atomicity, so if it
-- is present it is not in a corrupted state
unlessM (doesFileExist tarFP) $ do
req <- parseRequest $ T.unpack tarUrl
env <- asks scEnvAWS
let man = env ^. env_manager
withResponseUnliftIO req {decompress = const True} man $ \res -> do
throwErrorStatusCodes req res
createDirectoryIfMissing True $ takeDirectory tarFP
withBinaryFileDurableAtomic tarFP WriteMode $ \tarHandle ->
runConduitRes $ bodyReaderSource (responseBody res) .| sinkHandle tarHandle
void $ tryIO $ removeFile outname
withSystemTempDirectory ("hoogle-" ++ T.unpack (textDisplay snapName)) $ \tmpdir -> do
Any hasRestored <-
runConduitRes $
sourceFile tarFP .|
untar (restoreHoogleTxtFileWithCabal tmpdir snapshotId snapName) .|
foldMapC Any
-- We just check if we have any Hoogle .txt file at all.
unless hasRestored $ error "No Hoogle .txt files found"
-- Generate the hoogle database
let args = ["generate", "--database=" ++ outname, "--local=" ++ tmpdir]
logInfo $
mconcat
[ "Merging databases... ("
, foldMap fromString $ L.intersperse " " ("hoogle" : args)
, ")"
]
-- 'Hoogle.hoogle' expects to run as an app, and crashes if something
-- goes wrong. That's good.
liftIO $ Hoogle.hoogle args
logInfo "Merge done"
return $ Just outname
-- | Grabs hoogle txt file from the tarball and a matching cabal file from pantry. Writes
-- them into supplied temp directory and yields the result of operation as a boolean for
-- every tar entry.
restoreHoogleTxtFileWithCabal ::
FilePath
-> SnapshotId
-> SnapName
-> FileInfo
-> ConduitM ByteString Bool (ResourceT (RIO StackageCron)) ()
restoreHoogleTxtFileWithCabal tmpdir snapshotId snapName fileInfo =
case fileType fileInfo of
FTNormal -> do
let txtFileName = T.decodeUtf8With T.lenientDecode $ filePath fileInfo
txtPackageName = T.takeWhile (/= '.') txtFileName
mpkg = fromPathPiece txtPackageName
maybe (pure Nothing) (lift . lift . getSnapshotPackageCabalBlob snapshotId) mpkg >>= \case
Nothing -> do
logWarn $
"Unexpected hoogle filename: " <> display txtFileName <>
" in orig.tar for snapshot: " <>
display snapName
yield False
Just cabal -> do
writeFileBinary (tmpdir </> T.unpack txtPackageName <.> "cabal") cabal
sinkFile (tmpdir </> T.unpack txtFileName)
yield True
_ -> yield False
pathToPackageModule
:: Text
-- ^ Input is like @ace-0.6/ACE-Combinators@
-> Maybe (PackageIdentifierP, ModuleNameP)
pathToPackageModule txt =
case T.split (== '/') txt of
[pkgIdentifier, moduleNameDashes] -> do
modName :: ModuleNameP <- fromPathPiece moduleNameDashes
pkgId :: PackageIdentifierP <- fromPathPiece pkgIdentifier
Just (pkgId, modName)
_ -> Nothing