git-annex/Remote/Helper/Chunked.hs

410 lines
13 KiB
Haskell
Raw Normal View History

{- git-annex chunked remotes
-
- Copyright 2014 Joey Hess <id@joeyh.name>
-
- Licensed under the GNU GPL version 3 or higher.
-}
module Remote.Helper.Chunked (
ChunkSize,
ChunkConfig(..),
describeChunkConfig,
getChunkConfig,
storeChunks,
removeChunks,
retrieveChunks,
checkPresentChunks,
) where
import Common.Annex
import Utility.DataUnits
import Types.StoreRetrieve
import Types.Remote
import Types.Key
import Logs.Chunk
import Utility.Metered
import Crypto (EncKey)
import Backend (isStableKey)
import qualified Data.ByteString.Lazy as L
import qualified Data.Map as M
data ChunkConfig
= NoChunks
2014-07-24 19:08:07 +00:00
| UnpaddedChunks ChunkSize
| LegacyChunks ChunkSize
2014-08-01 19:36:11 +00:00
deriving (Show)
describeChunkConfig :: ChunkConfig -> String
describeChunkConfig NoChunks = "none"
describeChunkConfig (UnpaddedChunks sz) = describeChunkSize sz ++ "chunks"
describeChunkConfig (LegacyChunks sz) = describeChunkSize sz ++ " chunks (old style)"
describeChunkSize :: ChunkSize -> String
describeChunkSize sz = roughSize storageUnits False (fromIntegral sz)
noChunks :: ChunkConfig -> Bool
noChunks NoChunks = True
noChunks _ = False
getChunkConfig :: RemoteConfig -> ChunkConfig
getChunkConfig m =
case M.lookup "chunksize" m of
Nothing -> case M.lookup "chunk" m of
Nothing -> NoChunks
2014-08-01 19:36:11 +00:00
Just v -> readsz UnpaddedChunks v "chunk"
Just v -> readsz LegacyChunks v "chunksize"
where
2014-08-01 19:36:11 +00:00
readsz c v f = case readSize dataUnits v of
Just size
| size == 0 -> NoChunks
| size > 0 -> c (fromInteger size)
_ -> error $ "bad configuration " ++ f ++ "=" ++ v
-- An infinite stream of chunk keys, starting from chunk 1.
newtype ChunkKeyStream = ChunkKeyStream [Key]
chunkKeyStream :: Key -> ChunkSize -> ChunkKeyStream
chunkKeyStream basek chunksize = ChunkKeyStream $ map mk [1..]
where
mk chunknum = sizedk { keyChunkNum = Just chunknum }
sizedk = basek { keyChunkSize = Just (toInteger chunksize) }
nextChunkKeyStream :: ChunkKeyStream -> (Key, ChunkKeyStream)
nextChunkKeyStream (ChunkKeyStream (k:l)) = (k, ChunkKeyStream l)
nextChunkKeyStream (ChunkKeyStream []) = undefined -- stream is infinite!
takeChunkKeyStream :: ChunkCount -> ChunkKeyStream -> [Key]
takeChunkKeyStream n (ChunkKeyStream l) = genericTake n l
-- Number of chunks already consumed from the stream.
numChunks :: ChunkKeyStream -> Integer
numChunks = pred . fromJust . keyChunkNum . fst . nextChunkKeyStream
2014-07-26 16:04:35 +00:00
{- Splits up the key's content into chunks, passing each chunk to
- the storer action, along with a corresponding chunk key and a
- progress meter update callback.
-
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
- To support resuming, the checker is used to find the first missing
- chunk key. Storing starts from that chunk.
-
- This buffers each chunk in memory, so can use a lot of memory
- with a large ChunkSize.
- More optimal versions of this can be written, that rely
- on L.toChunks to split the lazy bytestring into chunks (typically
- smaller than the ChunkSize), and eg, write those chunks to a Handle.
- But this is the best that can be done with the storer interface that
- writes a whole L.ByteString at a time.
-}
storeChunks
:: UUID
-> ChunkConfig
-> Key
-> FilePath
-> MeterUpdate
-> Storer
-> CheckPresent
-> Annex Bool
storeChunks u chunkconfig k f p storer checker =
case chunkconfig of
(UnpaddedChunks chunksize) | isStableKey k ->
bracketIO open close (go chunksize)
_ -> storer k (FileContent f) p
where
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
open = tryIO $ openBinaryFile f ReadMode
close (Right h) = hClose h
close (Left _) = noop
go _ (Left e) = do
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
warning (show e)
return False
go chunksize (Right h) = do
let chunkkeys = chunkKeyStream k chunksize
(chunkkeys', startpos) <- seekResume h chunkkeys checker
b <- liftIO $ L.hGetContents h
gochunks p startpos chunksize b chunkkeys'
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
gochunks :: MeterUpdate -> BytesProcessed -> ChunkSize -> L.ByteString -> ChunkKeyStream -> Annex Bool
gochunks meterupdate startpos chunksize = loop startpos . splitchunk
where
splitchunk = L.splitAt chunksize
loop bytesprocessed (chunk, bs) chunkkeys
| L.null chunk && numchunks > 0 = do
-- Once all chunks are successfully
-- stored, update the chunk log.
chunksStored u k (FixedSizeChunks chunksize) numchunks
return True
| otherwise = do
liftIO $ meterupdate' zeroBytesProcessed
let (chunkkey, chunkkeys') = nextChunkKeyStream chunkkeys
ifM (storer chunkkey (ByteContent chunk) meterupdate')
( do
let bytesprocessed' = addBytesProcessed bytesprocessed (L.length chunk)
loop bytesprocessed' (splitchunk bs) chunkkeys'
, return False
)
where
numchunks = numChunks chunkkeys
{- The MeterUpdate that is passed to the action
- storing a chunk is offset, so that it reflects
- the total bytes that have already been stored
- in previous chunks. -}
meterupdate' = offsetMeterUpdate meterupdate bytesprocessed
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
{- Check if any of the chunk keys are present. If found, seek forward
- in the Handle, so it will be read starting at the first missing chunk.
- Returns the ChunkKeyStream truncated to start at the first missing
- chunk, and the number of bytes skipped due to resuming.
-
- As an optimisation, if the file fits into a single chunk, there's no need
- to check if that chunk is present -- we know it's not, because otherwise
- the whole file would be present and there would be no reason to try to
- store it.
-}
seekResume
:: Handle
-> ChunkKeyStream
-> CheckPresent
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
-> Annex (ChunkKeyStream, BytesProcessed)
seekResume h chunkkeys checker = do
sz <- liftIO (hFileSize h)
if sz <= fromMaybe 0 (keyChunkSize $ fst $ nextChunkKeyStream chunkkeys)
then return (chunkkeys, zeroBytesProcessed)
else check 0 chunkkeys sz
where
check pos cks sz
| pos >= sz = do
-- All chunks are already stored!
liftIO $ hSeek h AbsoluteSeek sz
return (cks, toBytesProcessed sz)
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
| otherwise = do
v <- tryNonAsync (checker k)
resume interrupted chunked uploads Leverage the new chunked remotes to automatically resume uploads. Sort of like rsync, although of course not as efficient since this needs to start at a chunk boundry. But, unlike rsync, this method will work for S3, WebDAV, external special remotes, etc, etc. Only directory special remotes so far, but many more soon! This implementation will also allow starting an upload from one repository, interrupting it, and then resuming the upload to the same remote from an entirely different repository. Note that I added a comment that storeKey should atomically move the content into place once it's all received. This was already an undocumented requirement -- it's necessary for hasKey to work reliably. This resume code just uses hasKey to find the first chunk that's missing. Note that if there are two uploads of the same key to the same chunked remote, one might resume at the point the other had gotten to, but both will then redundantly upload. As before. In the non-resume case, this adds one hasKey call per storeKey, and only if the remote is configured to use chunks. Future work: Try to eliminate that hasKey. Notice that eg, `git annex copy --to` checks if the key is present before sending it, so is already running hasKey.. which could perhaps be cached and reused. However, this additional overhead is not very large compared with transferring an entire large file, and the ability to resume is certianly worth it. There is an optimisation in place for small files, that avoids trying to resume if the whole file fits within one chunk. This commit was sponsored by Georg Bauer.
2014-07-28 18:18:08 +00:00
case v of
Right True ->
check pos' cks' sz
_ -> do
when (pos > 0) $
liftIO $ hSeek h AbsoluteSeek pos
return (cks, toBytesProcessed pos)
where
(k, cks') = nextChunkKeyStream cks
pos' = pos + fromMaybe 0 (keyChunkSize k)
{- Removes all chunks of a key from a remote, by calling a remover
2014-07-27 02:47:52 +00:00
- action on each.
-
- The remover action should succeed even if asked to
- remove a key that is not present on the remote.
-
- This action may be called on a chunked key. It will simply remove it.
-}
removeChunks :: (Key -> Annex Bool) -> UUID -> ChunkConfig -> EncKey -> Key -> Annex Bool
removeChunks remover u chunkconfig encryptor k = do
ls <- chunkKeys u chunkconfig k
2014-07-27 02:34:10 +00:00
ok <- allM (remover . encryptor) (concat ls)
when ok $ do
let chunksizes = catMaybes $ map (keyChunkSize <=< headMaybe) ls
forM_ chunksizes $ chunksRemoved u k . FixedSizeChunks . fromIntegral
return ok
{- Retrieves a key from a remote, using a retriever action.
-
- When the remote is chunked, tries each of the options returned by
- chunkKeys until it finds one where the retriever successfully
- gets the first chunked key. The content of that key, and any
- other chunks in the list is fed to the sink.
-
- If retrival of one of the subsequent chunks throws an exception,
- gives up and returns False. Note that partial data may have been
- written to the sink in this case.
-
- Resuming is supported when using chunks. When the destination file
- already exists, it skips to the next chunked key that would be needed
- to resume.
-}
retrieveChunks
:: Retriever
-> UUID
-> ChunkConfig
-> EncKey
-> Key
-> FilePath
-> MeterUpdate
-> (Maybe Handle -> Maybe MeterUpdate -> ContentSource -> Annex Bool)
-> Annex Bool
retrieveChunks retriever u chunkconfig encryptor basek dest basep sink
| noChunks chunkconfig =
-- Optimisation: Try the unchunked key first, to avoid
-- looking in the git-annex branch for chunk counts
-- that are likely not there.
getunchunked `catchNonAsync`
const (go =<< chunkKeysOnly u basek)
| otherwise = go =<< chunkKeys u chunkconfig basek
where
go ls = do
currsize <- liftIO $ catchMaybeIO $ getFileSize dest
let ls' = maybe ls (setupResume ls) currsize
2014-08-01 21:18:39 +00:00
if any null ls'
then return True -- dest is already complete
else firstavail currsize ls' `catchNonAsync` giveup
2014-07-27 03:01:44 +00:00
giveup e = do
warning (show e)
2014-07-27 03:01:44 +00:00
return False
firstavail _ [] = return False
firstavail currsize ([]:ls) = firstavail currsize ls
firstavail currsize ((k:ks):ls)
| k == basek = getunchunked
`catchNonAsync` (const $ firstavail currsize ls)
| otherwise = do
let offset = resumeOffset currsize k
let p = maybe basep
(offsetMeterUpdate basep . toBytesProcessed)
offset
v <- tryNonAsync $
retriever (encryptor k) p $ \content ->
bracketIO (maybe opennew openresume offset) hClose $ \h -> do
void $ tosink (Just h) p content
let sz = toBytesProcessed $
fromMaybe 0 $ keyChunkSize k
getrest p h sz sz ks
`catchNonAsync` giveup
case v of
Left e
| null ls -> giveup e
| otherwise -> firstavail currsize ls
Right r -> return r
getrest _ _ _ _ [] = return True
getrest p h sz bytesprocessed (k:ks) = do
let p' = offsetMeterUpdate p bytesprocessed
liftIO $ p' zeroBytesProcessed
ifM (retriever (encryptor k) p' $ tosink (Just h) p')
( getrest p h sz (addBytesProcessed bytesprocessed sz) ks
, giveup "chunk retrieval failed"
)
getunchunked = retriever (encryptor basek) basep $ tosink Nothing basep
opennew = openBinaryFile dest WriteMode
-- Open the file and seek to the start point in order to resume.
openresume startpoint = do
-- ReadWriteMode allows seeking; AppendMode does not.
h <- openBinaryFile dest ReadWriteMode
hSeek h AbsoluteSeek startpoint
return h
{- Progress meter updating is a bit tricky: If the Retriever
- populates a file, it is responsible for updating progress
- as the file is being retrieved.
-
- However, if the Retriever generates a lazy ByteString,
- it is not responsible for updating progress (often it cannot).
2014-08-03 05:21:38 +00:00
- Instead, the sink is passed a meter to update as it consumes
- the ByteString.
-}
tosink h p content = sink h p' content
where
p'
| isByteContent content = Just p
| otherwise = Nothing
{- Can resume when the chunk's offset is at or before the end of
- the dest file. -}
resumeOffset :: Maybe Integer -> Key -> Maybe Integer
resumeOffset Nothing _ = Nothing
resumeOffset currsize k
| offset <= currsize = offset
| otherwise = Nothing
where
offset = chunkKeyOffset k
{- Drops chunks that are already present in a file, based on its size.
- Keeps any non-chunk keys.
-}
setupResume :: [[Key]] -> Integer -> [[Key]]
setupResume ls currsize = map dropunneeded ls
where
dropunneeded [] = []
dropunneeded l@(k:_) = case keyChunkSize k of
Just chunksize | chunksize > 0 ->
genericDrop (currsize `div` chunksize) l
_ -> l
{- Checks if a key is present in a remote. This requires any one
- of the lists of options returned by chunkKeys to all check out
- as being present using the checker action.
-
- Throws an exception if the remote is not accessible.
-}
checkPresentChunks
:: CheckPresent
-> UUID
-> ChunkConfig
-> EncKey
-> Key
-> Annex Bool
checkPresentChunks checker u chunkconfig encryptor basek
| noChunks chunkconfig = do
-- Optimisation: Try the unchunked key first, to avoid
-- looking in the git-annex branch for chunk counts
-- that are likely not there.
v <- check basek
case v of
Right True -> return True
Left e -> checklists (Just e) =<< chunkKeysOnly u basek
_ -> checklists Nothing =<< chunkKeysOnly u basek
2014-07-29 19:07:32 +00:00
| otherwise = checklists Nothing =<< chunkKeys u chunkconfig basek
where
checklists Nothing [] = return False
checklists (Just deferrederror) [] = throwM deferrederror
2014-07-29 19:07:32 +00:00
checklists d (l:ls)
| not (null l) = do
v <- checkchunks l
case v of
2014-07-29 19:07:32 +00:00
Left e -> checklists (Just e) ls
Right True -> return True
2014-07-29 19:07:32 +00:00
Right False -> checklists Nothing ls
| otherwise = checklists d ls
checkchunks :: [Key] -> Annex (Either SomeException Bool)
checkchunks [] = return (Right True)
checkchunks (k:ks) = do
v <- check k
case v of
Right True -> checkchunks ks
Right False -> return $ Right False
Left e -> return $ Left e
check = tryNonAsync . checker . encryptor
2014-07-27 05:24:34 +00:00
{- A key can be stored in a remote unchunked, or as a list of chunked keys.
- This can be the case whether or not the remote is currently configured
- to use chunking.
-
- It's even possible for a remote to have the same key stored multiple
- times with different chunk sizes!
-
- This finds all possible lists of keys that might be on the remote that
- can be combined to get back the requested key, in order from most to
- least likely to exist.
2014-07-27 05:24:34 +00:00
-}
chunkKeys :: UUID -> ChunkConfig -> Key -> Annex [[Key]]
chunkKeys u chunkconfig k = do
l <- chunkKeysOnly u k
return $ if noChunks chunkconfig
then [k] : l
else l ++ [[k]]
chunkKeysOnly :: UUID -> Key -> Annex [[Key]]
chunkKeysOnly u k = map (toChunkList k) <$> getCurrentChunks u k
2014-07-27 05:24:34 +00:00
toChunkList :: Key -> (ChunkMethod, ChunkCount) -> [Key]
toChunkList k (FixedSizeChunks chunksize, chunkcount) =
takeChunkKeyStream chunkcount $ chunkKeyStream k chunksize
toChunkList _ (UnknownChunks _, _) = []