mirror of
https://github.com/hasura/graphql-engine.git
synced 2024-12-17 20:41:49 +03:00
7bead93827
Query plan caching was introduced by - I believe - hasura/graphql-engine#1934 in order to reduce the query response latency. During the development of PDV in hasura/graphql-engine#4111, it was found out that the new architecture (for which query plan caching wasn't implemented) performed comparably to the pre-PDV architecture with caching. Hence, it was decided to leave query plan caching until some day in the future when it was deemed necessary. Well, we're in the future now, and there still isn't a convincing argument for query plan caching. So the time has come to remove some references to query plan caching from the codebase. For the most part, any code being removed would probably not be very well suited to the post-PDV architecture of query execution, so arguably not much is lost. Apart from simplifying the code, this PR will contribute towards making the GraphQL schema generation more modular, testable, and easier to profile. I'd like to eventually work towards a situation in which it's easy to generate a GraphQL schema parser *in isolation*, without being connected to a database, and then parse a GraphQL query *in isolation*, without even listening any HTTP port. It is important that both of these operations can be examined in detail, and in isolation, since they are two major performance bottlenecks, as well as phases where many important upcoming features hook into. Implementation The following have been removed: - The entirety of `server/src-lib/Hasura/GraphQL/Execute/Plan.hs` - The core phases of query parsing and execution no longer have any references to query plan caching. Note that this is not to be confused with query *response* caching, which is not affected by this PR. This includes removal of the types: - - `Opaque`, which is replaced by a tuple. Note that the old implementation was broken and did not adequately hide the constructors. - - `QueryReusability` (and the `markNotReusable` method). Notably, the implementation of the `ParseT` monad now consists of two, rather than three, monad transformers. - Cache-related tests (in `server/src-test/Hasura/CacheBoundedSpec.hs`) have been removed . - References to query plan caching in the documentation. - The `planCacheOptions` in the `TenantConfig` type class was removed. However, during parsing, unrecognized fields in the YAML config get ignored, so this does not cause a breaking change. (Confirmed manually, as well as in consultation with @sordina.) - The metrics no longer send cache hit/miss messages. There are a few places in which one can still find references to query plan caching: - We still accept the `--query-plan-cache-size` command-line option for backwards compatibility. The `HASURA_QUERY_PLAN_CACHE_SIZE` environment variable is not read. https://github.com/hasura/graphql-engine-mono/pull/1815 GitOrigin-RevId: 17d92b254ec093c62a7dfeec478658ede0813eb7
187 lines
6.5 KiB
Haskell
187 lines
6.5 KiB
Haskell
{-# LANGUAGE DuplicateRecordFields #-}
|
|
|
|
{-|
|
|
Counters used in telemetry collection. Additional counters can be added here.and
|
|
serviced in "Hasura.Server.Telemetry".
|
|
-}
|
|
|
|
module Hasura.Server.Telemetry.Counters
|
|
(
|
|
-- * Service timing and counts, by various dimensions
|
|
-- ** Local metric recording
|
|
recordTimingMetric
|
|
, RequestDimensions(..), RequestTimings(..)
|
|
-- *** Dimensions
|
|
, QueryType(..), Locality(..), Transport(..)
|
|
-- ** Metric upload
|
|
, dumpServiceTimingMetrics
|
|
, ServiceTimingMetrics(..)
|
|
, ServiceTimingMetric(..)
|
|
, RunningTimeBucket(..)
|
|
, RequestTimingsCount(..)
|
|
)
|
|
where
|
|
|
|
import Hasura.Prelude
|
|
|
|
import qualified Data.Aeson as A
|
|
import qualified Data.Aeson.TH as A
|
|
import qualified Data.HashMap.Strict as HM
|
|
|
|
import Data.IORef
|
|
import Data.Time.Clock.POSIX (POSIXTime, getPOSIXTime)
|
|
import GHC.IO.Unsafe (unsafePerformIO)
|
|
|
|
|
|
-- | The properties that characterize this request. The dimensions over which
|
|
-- we collect metrics for each serviced request.
|
|
data RequestDimensions =
|
|
RequestDimensions
|
|
{ telemQueryType :: !QueryType
|
|
, telemLocality :: !Locality
|
|
, telemTransport :: !Transport
|
|
}
|
|
deriving (Show, Generic, Eq)
|
|
|
|
instance Hashable RequestDimensions
|
|
|
|
|
|
-- | Accumulated time metrics.
|
|
data RequestTimings =
|
|
RequestTimings {
|
|
telemTimeIO :: !Seconds
|
|
-- ^ Time spent waiting on PG/remote http calls
|
|
, telemTimeTot :: !Seconds
|
|
-- ^ Total service time for request (including 'telemTimeIO')
|
|
}
|
|
|
|
-- | Sum
|
|
instance Semigroup RequestTimings where
|
|
RequestTimings a b <> RequestTimings x y = RequestTimings (a+x) (b+y)
|
|
|
|
-- | 'RequestTimings' along with the count
|
|
data RequestTimingsCount =
|
|
RequestTimingsCount {
|
|
telemTimeIO :: !Seconds
|
|
, telemTimeTot :: !Seconds
|
|
, telemCount :: !Word
|
|
-- ^ The number of requests that have contributed to the accumulated timings above.
|
|
-- So e.g. @telemTimeTot / count@ would give the mean service time.
|
|
}
|
|
deriving (Show, Generic, Eq)
|
|
|
|
-- | Sum
|
|
instance Semigroup RequestTimingsCount where
|
|
RequestTimingsCount a b c <> RequestTimingsCount x y z =
|
|
RequestTimingsCount (a+x) (b+y) (c+z)
|
|
|
|
-- | Internal. Counts and durations across many 'RequestDimensions'.
|
|
--
|
|
-- NOTE: We use the global mutable variable pattern for metric collection
|
|
-- counters for convenience at collection site (don't wear hairshirts that
|
|
-- discourage useful reporting).
|
|
requestCounters :: IORef (HM.HashMap (RequestDimensions, RunningTimeBucket) RequestTimingsCount)
|
|
{-# NOINLINE requestCounters #-}
|
|
requestCounters = unsafePerformIO $ newIORef HM.empty
|
|
|
|
-- | Internal. Since these metrics are accumulated while graphql-engine is
|
|
-- running and sent periodically, we need to include a tag that is unique for
|
|
-- each start of hge. This lets us e.g. query for just the latest uploaded
|
|
-- sample for each start of hge.
|
|
--
|
|
-- We use time rather than a UUID since having this be monotonic increasing is
|
|
-- convenient.
|
|
approxStartTime :: POSIXTime
|
|
{-# NOINLINE approxStartTime #-}
|
|
approxStartTime = unsafePerformIO getPOSIXTime
|
|
|
|
-- | Was this request a mutation (involved DB writes)?
|
|
data QueryType = Mutation | Query
|
|
deriving (Enum, Show, Eq, Generic)
|
|
instance Hashable QueryType
|
|
instance A.ToJSON QueryType
|
|
instance A.FromJSON QueryType
|
|
|
|
-- | Was this a PG local query, or did it involve remote execution?
|
|
data Locality
|
|
= Empty -- ^ No data was fetched
|
|
| Local -- ^ local DB data
|
|
| Remote -- ^ remote schema
|
|
| Heterogeneous -- ^ mixed
|
|
deriving (Enum, Show, Eq, Generic)
|
|
instance Hashable Locality
|
|
instance A.ToJSON Locality
|
|
instance A.FromJSON Locality
|
|
instance Semigroup Locality where
|
|
Empty <> x = x
|
|
x <> Empty = x
|
|
x <> y | x == y = x
|
|
_ <> _ = Heterogeneous
|
|
instance Monoid Locality where
|
|
mempty = Empty
|
|
|
|
-- | Was this a query over http or websockets?
|
|
data Transport = HTTP | WebSocket
|
|
deriving (Enum, Show, Eq, Generic)
|
|
instance Hashable Transport
|
|
instance A.ToJSON Transport
|
|
instance A.FromJSON Transport
|
|
|
|
-- | The timings and counts here were from requests with total time longer than
|
|
-- 'bucketGreaterThan' (but less than any larger bucket cutoff times).
|
|
newtype RunningTimeBucket = RunningTimeBucket { bucketGreaterThan :: Seconds }
|
|
deriving (Ord, Eq, Show, Generic, A.ToJSON, A.FromJSON, Hashable)
|
|
|
|
|
|
-- NOTE: an HDR histogram is a nice way to collect metrics when you don't know
|
|
-- a priori what the most useful binning is. It's not clear how we'd make use
|
|
-- of that here though. So these buckets are arbitrary, and can be adjusted as
|
|
-- needed, but we shouldn't have more than a handful to keep payload size down.
|
|
totalTimeBuckets :: [RunningTimeBucket]
|
|
totalTimeBuckets = coerce [0.000, 0.001, 0.050, 1.000, 3600.000 :: Seconds]
|
|
|
|
-- | Save a timing metric sample in our in-memory store. These will be
|
|
-- accumulated and uploaded periodically in "Hasura.Server.Telemetry".
|
|
recordTimingMetric :: MonadIO m => RequestDimensions -> RequestTimings -> m ()
|
|
recordTimingMetric reqDimensions RequestTimings{..} = liftIO $ do
|
|
let ourBucket = fromMaybe (RunningTimeBucket 0) $ -- although we expect 'head' would be safe here
|
|
listToMaybe $ dropWhile (> coerce telemTimeTot) $
|
|
reverse $ sort totalTimeBuckets
|
|
atomicModifyIORef' requestCounters $ (,()) .
|
|
HM.insertWith (<>) (reqDimensions, ourBucket) RequestTimingsCount{telemCount = 1, ..}
|
|
|
|
-- | The final shape of this part of our metrics data JSON. This should allow
|
|
-- reasonably efficient querying using GIN indexes and JSONB containment
|
|
-- operations (which treat arrays as sets).
|
|
data ServiceTimingMetrics
|
|
= ServiceTimingMetrics
|
|
{ collectionTag :: Int
|
|
-- ^ This is set to a new unique value when the counters reset (e.g. because of a restart)
|
|
, serviceTimingMetrics :: [ServiceTimingMetric]
|
|
}
|
|
deriving (Show, Generic, Eq)
|
|
data ServiceTimingMetric
|
|
= ServiceTimingMetric
|
|
{ dimensions :: RequestDimensions
|
|
, bucket :: RunningTimeBucket
|
|
, metrics :: RequestTimingsCount
|
|
}
|
|
deriving (Show, Generic, Eq)
|
|
|
|
|
|
$(A.deriveJSON hasuraJSON ''RequestTimingsCount)
|
|
$(A.deriveJSON hasuraJSON ''RequestDimensions)
|
|
|
|
instance A.ToJSON ServiceTimingMetric
|
|
instance A.FromJSON ServiceTimingMetric
|
|
instance A.ToJSON ServiceTimingMetrics
|
|
instance A.FromJSON ServiceTimingMetrics
|
|
|
|
dumpServiceTimingMetrics :: MonadIO m => m ServiceTimingMetrics
|
|
dumpServiceTimingMetrics = liftIO $ do
|
|
cs <- readIORef requestCounters
|
|
let serviceTimingMetrics = flip map (HM.toList cs) $
|
|
\((dimensions, bucket), metrics)-> ServiceTimingMetric{..}
|
|
collectionTag = round approxStartTime
|
|
return ServiceTimingMetrics{..}
|