Status
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
This FLIP is initially proposed by Yuan Zhu (zstraw@163.com), and finished by Qingsheng Ren (partial caching part) and Alexander Smirnov (full caching part).
Motivation
As a widely-used feature in Flink SQL jobs, the performance of lookup table source is essential not only for users but also source developers for tuning their implementations. Most lookup table sources use cache to achieve better performance, but there are some features missing in the current design of cache:
- Missing cache related metrics, which is the key to debug and optimize SQL tasks
- Duplicated implementations. Currently every lookup source needs to implement or use its own cache.
- Inconsistent. Table options related to caching are defined differently in sources
In order to address the issues above, we propose here to define a unified abstraction for lookup source cache and its related metrics.
Proposed Changes
Top-level APIs
In order to clarify the semantic of lookup, we'd like to introduce some top-level APIs for general lookup operations without caching:
LookupFunction
/AsyncLookupFunction
, an extended version of TableFunction to make the API more straight forward.LookupFunctionProvider
/AsyncLookupProvider
, serve as the creator of LookupFunction / AsyncLookupFunction in table source
And APIs related to the cache:
LookupCache
, defining the cache used in lookup table.DefaultLookupCache
a default implementation of a cache that suitable for most use cases.CacheMetricGroup
, defining metrics should be reported by the cache.
Partial and Full Caching
More specifically, we'd like to provide public interfaces for the most 2 common cases to lookup source developers, which are named as partial and full caching.
Partial caching
Partial caching loads data into the cache along with the access to the external system. If the key to lookup does not exist in the cache, a lookup action to the external system will be triggered and the lookup result will be stored in the cache for further loading. Users and lookup table developers are able to config the eviction policy and maximum size of the cache.
In order to support partial caching, we propose to introduce 2 new interfaces:
PartialCachingLookupProvider
/AsyncPartialCachingLookupProvider
, as the API interacting with table source to get LookupFunction and LookupCache.
The cache serves as a component in LookupJoinRunner, and would be pluggable by specifying in the constructor of the provider. The planner will take over the lookup function and the cache created from the provider and pass it to the LookupJoinRunner. The cache will be instantiated during the runtime execution and loading operations via lookup function if there's a cache miss.
Full Caching
If the size of lookup table is relatively small to fit into the memory, and the lookup table doesn't change frequently, it'll be more efficient to load all entries of the lookup table into the cache to reduce network I/O, and refresh the table periodically. We'd like to name this use case as "full cache". Logically the reload operation is a kind of scan, so we'd like to reuse the ScanRuntimeProvider so that developers could reuse the scanning logic implemented in Source / SourceFunction / InputFormat. Considering the complexity of Source API, we'd like to support SourceFunction and InputFormat API first. Supporting Source API might require new topology and will be discussed later in another FLIP.
We propose to introduce several new interfaces:
FullCachingLookupProvider
, for reusing the ability of scanning.CacheReloadTrigger
, for customizing reloading strategies of all entries in the full cache.
Also we'd like to provide two default implementations of CacheReloadTrigger:
PeriodicCacheReloadTrigger
, for triggering reload periodically with a specific interval- TimedCacheReloadTrigger, for triggering reload at the specific time and repeat with the interval in days.
Public Interfaces
Lookup Functions
As the usage of TableFunction interface is not quite straight forward to lookup table developers, we'd like to introduce a new interface for sync and async lookup tables. Caching will be only supported on LookupFunction / AsyncLookupFunction.
/** * A wrapper class of {@link TableFunction} for synchronously lookup rows matching the lookup keys * from external system. * * <p>The output type of this table function is fixed as {@link RowData}. */ @PublicEvolving public abstract class LookupFunction extends TableFunction<RowData> { /** * Synchronously lookup rows matching the lookup keys. * * @param keyRow - A {@link RowData} that wraps keys to lookup. * @return A collections of all matching rows in the lookup table. */ public abstract Collection<RowData> lookup(RowData keyRow) throws IOException; /** Invoke {@link #lookup} and handle exceptions. */ public final void eval(Object... keys) { try { lookup(GenericRowData.of(keys)).forEach(this::collect); } catch (IOException e) { throw new RuntimeException("Failed to lookup values with given key", e); } } }
/** * A wrapper class of {@link AsyncTableFunction} for asynchronously lookup rows matching the lookup * keys from external system. * * <p>The output type of this table function is fixed as {@link RowData}. */ @PublicEvolving public abstract class AsyncLookupFunction extends AsyncTableFunction<RowData> { /** * Asynchronously lookup rows matching the lookup keys. * * @param keyRow - A {@link RowData} that wraps keys to lookup. * @return A collections of all matching rows in the lookup table. */ public abstract CompletableFuture<Collection<RowData>> asyncLookup(RowData keyRow); /** Invokes {@link #asyncLookup} and chains futures. */ public final void eval(CompletableFuture<Collection<RowData>> future, Object... keys) { asyncLookup(GenericRowData.of(keys)) .whenCompleteAsync( (result, exception) -> { if (exception != null) { future.completeExceptionally(exception); return; } future.complete(result); }); } }
LookupFunctionProvider
/** A provider for creating {@link LookupFunction}. */ @PublicEvolving public interface LookupFunctionProvider extends LookupTableSource.LookupRuntimeProvider { static LookupFunctionProvider of(LookupFunction lookupFunction) { return () -> lookupFunction; } LookupFunction createLookupFunction(); }
AsyncLookupFunctionProvider
/** A provider for creating {@link AsyncLookupFunction}. */ @PublicEvolving public interface AsyncLookupFunctionProvider extends LookupTableSource.LookupRuntimeProvider { static AsyncLookupFunctionProvider of(AsyncLookupFunction asyncLookupFunction) { return () -> asyncLookupFunction; } AsyncLookupFunction createAsyncLookupFunction(); }
LookupCache
Considering there might be custom caching strategies and optimizations, we'd like to expose the cache interface as public API for developers to make the cache pluggable.
/** * A semi-persistent mapping from keys to values for storing entries of lookup table. * * <p>The type of the caching key is a {@link RowData} with lookup key fields packed inside. The * type of value is a {@link Collection} of {@link RowData}, which are rows matching lookup key * fields. * * <p>Cache entries are manually added using {@link #put}, and are stored in the cache until either * evicted or manually invalidated. * * <p>Implementations of this interface are expected to be thread-safe, and can be safely accessed * by multiple concurrent threads. */ @PublicEvolving public interface LookupCache extends AutoClosable, Serializable { /** * Initialize the cache. * * @param metricGroup the metric group to register cache related metrics. */ void open(CacheMetricGroup metricGroup); /** * Returns the value associated with key in this cache, or null if there is no cached value for * key. */ @Nullable Collection<RowData> getIfPresent(RowData key); /** * Associates the specified value rows with the specified key row in the cache. If the cache * previously contained value associated with the key, the old value is replaced by the * specified value. * * @return the previous value rows associated with key, or null if there was no mapping for key. * @param key - key row with which the specified value is to be associated * @param value – value rows to be associated with the specified key */ Collection<RowData> put(RowData key, Collection<RowData> value); /** Discards any cached value for the specified key. */ void invalidate(RowData key); /** Returns the number of key-value mappings in the cache. */ long size(); }
DefaultLookupCache
As the cache should be instantiated during runtime execution to avoid serialization / deserialization, a factory is required for creating the cache.
/** Default implementation of {@link LookupCache}. */ @PublicEvolving public class DefaultLookupCache implements LookupCache { private final Duration expireAfterAccessDuration; private final Duration expireAfterWriteDuration; private final Long maximumSize; private final boolean cacheMissingKey; private DefaultLookupCache( Duration expireAfterAccessDuration, Duration expireAfterWriteDuration, Long maximumSize, boolean cacheMissingKey) { this.expireAfterAccessDuration = expireAfterAccessDuration; this.expireAfterWriteDuration = expireAfterWriteDuration; this.initialCapacity = initialCapacity; this.maximumSize = maximumSize; this.cacheMissingKey = cacheMissingKey; } public static Builder newBuilder() { return new Builder(); } public static class Builder { private Duration expireAfterAccessDuration; private Duration expireAfterWriteDuration; private Long maximumSize; private Boolean cacheMissingKey; public Builder expireAfterAccess(Duration duration) { expireAfterAccessDuration = duration; return this; } public Builder expireAfterWrite(Duration duration) { expireAfterWriteDuration = duration; return this; } public Builder maximumSize(long maximumSize) { this.maximumSize = maximumSize; return this; } public Builder cacheMissingKey(boolean cacheMissingKey) { this.cacheMissingKey = cacheMissingKey; return this; } public DefaultLookupCache build() { return new DefaultLookupCache( expireAfterAccessDuration, expireAfterWriteDuration, maximumSize, cacheMissingKey); } } }
CacheMetricGroup
An interface defining all cache related metric:
/** * Pre-defined metrics for cache. * * <p>Please note that these methods should only be invoked once. Registering a metric with same * name for multiple times would lead to an undefined behavior. */ @PublicEvolving public interface CacheMetricGroup extends MetricGroup { /** The number of cache hits. */ void hitCounter(Counter hitCounter); /** The number of cache misses. */ void missCounter(Counter missCounter); /** The number of times to load data into cache from external system. */ void loadCounter(Counter loadCounter); /** The number of load failures. */ void numLoadFailuresCounter(Counter numLoadFailuresCounter); /** The time spent for the latest load operation. */ void latestLoadTimeGauge(Gauge<Long> latestLoadTimeGauge); /** The number of records in cache. */ void numCachedRecordsGauge(Gauge<Long> numCachedRecordsGauge); /** The number of bytes used by cache. */ void numCachedBytesGauge(Gauge<Long> numCachedBytesGauge); }
PartialCachingLookupProvider
This is the API between table framework and user's table source. Implementation should define how to create a lookup function and whether to use cache.
/** * Provider for creating {@link LookupFunction} and {@link LookupCache} for storing lookup entries. */ @PublicEvolving public interface PartialCachingLookupProvider extends LookupFunctionProvider { /** * Build a {@link PartialCachingLookupProvider} from the specified {@link LookupFunction} and * {@link LookupCache}. */ static PartialCachingLookupProvider of(LookupFunction lookupFunction, LookupCache cache) { return new PartialCachingLookupProvider() { @Override public LookupCache getCache() { return cache; } @Override public LookupFunction createLookupFunction() { return lookupFunction; } }; } /** Get a new instance of {@link LookupCache}. */ LookupCache getCache(); }
PartialCachingAsyncLookupProvider
/** * Provider for creating {@link AsyncLookupFunction} and {@link LookupCache} for storing lookup entries. */ @PublicEvolving public interface PartialCachingAsyncLookupProvider extends AsyncLookupFunctionProvider { /** * Build a {@link PartialCachingLookupProvider} from the specified {@link AsyncLookupFunction} and * {@link LookupCache}. */ static PartialCachingLookupProvider of(AsyncLookupFunction asyncLookupFunction, LookupCache cache) { return new PartialCachingAsyncLookupProvider() { @Override public LookupCache getCache() { return cache; } @Override public AsyncLookupFunction createAsyncLookupFunction() { return asyncLookupFunction; } }; } /** Get a new instance of {@link LookupCache}. */ LookupCache getCache(); }
FullCachingLookupProvider
This interface is for supporting full cache strategy. It reuses ScanRuntimeProvider and defines reload time.
/** * A {@link CachingLookupProvider} that never lookup in external system on cache miss and provides a * cache for holding all entries in the external system. The cache will be fully reloaded from the * external system by the {@link ScanTableSource.ScanRuntimeProvider} and reload operations will be * triggered by the {@link CacheReloadTrigger}. */ @PublicEvolving public interface FullCachingLookupProvider extends LookupFunctionProvider { static FullCachingLookupProvider of( ScanTableSource.ScanRuntimeProvider scanRuntimeProvider, CacheReloadTrigger cacheReloadTrigger) { return new FullCachingLookupProvider() { @Override public ScanTableSource.ScanRuntimeProvider getScanRuntimeProvider() { return scanRuntimeProvider; } @Override public CacheReloadTrigger getCacheReloadTrigger() { return cacheReloadTrigger; } @Override public LookupFunction createLookupFunction() { return keyRow -> null; } }; } /** * Get a {@link ScanTableSource.ScanRuntimeProvider} for scanning all entries from the external * lookup table and load into the cache. */ ScanTableSource.ScanRuntimeProvider getScanRuntimeProvider(); /** Get a {@link CacheReloadTrigger} for triggering the reload operation. */ CacheReloadTrigger getCacheReloadTrigger(); }
CacheReloadTrigger
A trigger defining custom logic for triggering full cache reloading.
/** Customized trigger for reloading all lookup table entries in full caching mode. */ @PublicEvolving public interface CachingReloadTrigger extends AutoCloseable, Serializable { /** Open the trigger. */ void open(Context context) throws Exception; /** * Context of {@link CacheReloadTrigger} for getting information about times and * triggering reload. */ interface Context { /** Get current processing time. */ long currentProcessingTime(); /** Get current watermark on the main stream. */ long currentWatermark(); /** Trigger a reload operation on the full cache. */ CompletableFuture<Void> triggerReload(); } }
PeriodicCacheReloadTrigger
An implementation of FullCachingReloadTrigger that triggers reload with a specified interval.
/** A trigger that reloads all entries periodically with specified interval or delay. */ public class PeriodicCacheReloadTrigger implements CacheReloadTrigger { private final Duration reloadInterval; private final ScheduleMode scheduleMode; public PeriodicCacheReloadTrigger(Duration reloadInterval, ScheduleMode scheduleMode) { this.reloadInterval = reloadInterval; this.scheduleMode = scheduleMode; } @Override public void open(CacheReloadTrigger.Context context) { // Register periodic reload task } @Override public void close() throws Exception { // Dispose resources } public enum ScheduleMode { FIXED_DELAY, FIXED_RATE } }
TimedCacheReloadTrigger
/** A trigger that reloads at a specific local time and repeat for the given interval in days. */ public class TimedCacheReloadTrigger implements CacheReloadTrigger { private final LocalTime reloadTime; private final int reloadIntervalInDays; public TimedCacheReloadTrigger(LocalTime reloadTime, int reloadIntervalInDays) { this.reloadTime = reloadTime; this.reloadIntervalInDays = reloadIntervalInDays; } @Override public void open(Context context) { // Register periodic reload task } @Override public void close() throws Exception { // Dispose resources } }
TableFunctionProvider / AsyncTableFunctionProvider
We'd like to deprecate these two interfaces and let developers to switch to the new LookupFunctionProvider / AsyncLookupFunctionProvider / FullCachingLookupProvider instead.
Table Options for Lookup Cache
In order to unify the usage of caching across all connectors, we'd like to introduce some common table options, which are defined under class LookupOptions
. Note that these options are not required to implement by all connectors.
Option | Type | Descriptions |
---|---|---|
lookup.cache | Enum of NONE, PARTIAL and FULL | The caching strategy for this lookup table. NONE: Do not use cache Partial: Use partial caching mode FULL: Use full caching mode |
lookup.max-retries | Integer | The maximum allowed retries if a lookup operation fails |
lookup.partial-cache.expire-after-access | Duration | Duration to expire an entry in the cache after accessing |
lookup.partial-cache.expire-after-write | Duration | Duration to expire an entry in the cache after writing |
lookup.partial-cache.cache-missing-key | Boolean | Whether to store an empty value into the cache if the lookup key doesn't match any rows in the table |
lookup.partial-cache.max-rows | Long | The maximum number of rows to store in the cache |
lookup.full-cache.reload-strategy | Enum of PERIODIC and TIMED | The reload strategy for the full cache scenario. PERIODIC: Use PeriodicCacheReloadTrigger TIMED: Use TimedCacheReloadTrigger |
lookup.full-cache.periodic-reload.interval | Duration | Duration to trigger reload in the PeriodicCacheReloadTrigger |
lookup.full-cache.periodic-reload.schedule-mode | Enum of FIXED_DELAY and FIXED_RATE | The periodically schedule mode of reloading in the PeriodicCacheReloadTrigger |
lookup.full-cache.timed-reload.iso-time | String | Time in ISO-8601 format when cache needs to be reloaded. Time can be specified either with timezone or without timezone (target JVM local timezone will be used). See formatter ISO_TIME. |
lookup.full-cache.timed-reload.interval-in-days | Integer | The interval in days to trigger the reload at the specified time |
Cache Metrics
It is important to mention that a cache implementation does not have to report all the defined metrics. But if a cache reports a metric of the same semantic defined below, the implementation should follow the convention.
Name | Type | Unit | Description |
---|---|---|---|
numCachedRecord | Gauge | Records | The number of records in cache. |
numCachedBytes | Gauge | Bytes | The number of bytes used by cache. |
hitCount | Counter | The number of cache hits | |
missCount | Counter | The number of cache misses, which might leads to loading operations | |
loadCount | Counter | The number of times to load data into cache from external system. For partial cache the load count should be equal to miss count, but for all cache this would be different. | |
numLoadFailure | Counter | The number of load failures | |
latestLoadTime | Gauge | ms | The time spent for the latest load operation |
Here we just define fundamental metrics and let the external metric system make the aggregation to get more descriptive values such as hitRate = hitCount / (hitCount + missCount).
Scope
The metric group for the cache would be a sub-group of the OperatorMetricGroup where the table function belongs to.
Future Works
In order to reduce network I/O with external systems and the usage of cache further, some optimizations implemented on scan source could be also applied on the lookup table, such as projection and filter pushdown. These features will be introduced separately in another FLIP.
Compatibility, Deprecation, and Migration Plan
Currently we have JDBC, Hive and HBase connector implemented lookup table source. All existing implementations will be migrated to the current design and the migration will be transparent to end users. Table options related to caching defined by these connectors will be migrated to new table options defined in this FLIP above.
Test Plan
We will use unit and integration test for validating the functionality of cache implementations.
Rejected Alternatives
Add cache in TableFunction implementations
Compared with this design, adding cache in TableFunction implementations might lead to inconsistency between sync and async table function, and not suitable for applying optimizations.