feat: Verification with sampling #28

mwalkiewicz · 2021-09-29T10:44:22Z

This change is

prawilny

Looks good to me, but I left some remarks - nitpicks mostly. Clicked "approve" so you can merge if you consider them unimportant.

Reviewed 17 of 17 files at r1, all commit messages.
Reviewable status: all files reviewed, 3 unresolved discussions (waiting on @dopiera and @mwalkiewicz)

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/readsampling/PartialReadSamplingStrategy.java, line 21 at r1 (raw file):

  private final float interval;
  private float counter;

It seems to be working but in general I like my counters to be integers.
@dopiera, what are the assumptions - I mean: what percentage of verified reads do we expect to deal with - are here?

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/test/java/com/google/cloud/bigtable/mirroring/hbase1_x/Helpers.java, line 34 at r1 (raw file):

import org.apache.hadoop.hbase.client.Result;

public class Helpers {

Please create an issue after merging for someone (probably me) to use this package in 2.x client tests.

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/test/java/com/google/cloud/bigtable/mirroring/hbase1_x/TestVerificationSampling.java, line 60 at r1 (raw file):

@RunWith(JUnit4.class)
public class TestVerificationSampling {

Maybe a single case each for a write, a readless batch and an error on primary database?

dopiera

Reviewed 17 of 17 files at r1, all commit messages.
Reviewable status: all files reviewed, 7 unresolved discussions (waiting on @mwalkiewicz and @prawilny)

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/MirroringTable.java, line 122 at r1 (raw file):

      SecondaryWriteErrorConsumerWithMetrics secondaryWriteErrorConsumer,
      MirroringTracer mirroringTracer,
      ReadSamplingStrategy readSamplingStrategy) {

I'd reorder the last two parameters. My rationale is importance - tracer is only for monitoring but sampling actually affect the table's behavior.

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/BatchHelpers.java, line 190 at r1 (raw file):

    }

    public SplitBatchResponse(

I think it would be more readable to not change this class, but change whoever consumes its work.

I understand that it is appealing to do it the way you did because it's only one place, but I think this class is already complicated enough. Effectively, this class is a pretty version of a function with 3 inputs and 8 outputs whose explanations are already complicated (some have nulls and retain indices, some are lists). That way, we'd change the explanation of 4 of the outputs.

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/MirroringConfigurationHelper.java, line 104 at r1 (raw file):

   * inclusive. 0 disables the verification completely, 1 enables verification of every read.
   */
  public static final String MIRRORING_READ_VERIFICATION_RATE =

PTAL at my further comment. I'd make it an integer denoting a percentage.

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/readsampling/PartialReadSamplingStrategy.java, line 18 at r1 (raw file):

package com.google.cloud.bigtable.mirroring.hbase1_x.utils.readsampling;

public class PartialReadSamplingStrategy implements ReadSamplingStrategy {

I think I'd change two things:

make it probabilistic - the cost of generating a pseudo-random number vs whole request processing is negligible, but you are much less likely to deterministically omit some operations
get rid of the floats in the implementation and interface (so that if someone puts 0 or 100 it is really 0 or 100 - something you cannot guarantee with floats)
ditch the abstraction altogether and hardcode a simple implementation (suggested below) - I don't believe anyone would need any more sophisticated algorithm in our mirroring client

Here's how I'd implement it:

class RandomSample {
  boolean sample() {
    return ThreadLocalRandom.current().nextInt(100) < probPerc;
  }
  int probPerc;  // some configurable number [0..100]
};

This way you don't need float arithmetic or special cases for 0.

WDYT?

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/readsampling/PartialReadSamplingStrategy.java, line 21 at r1 (raw file):

Previously, prawilny (Adam Czajkowski) wrote…

It seems to be working but in general I like my counters to be integers.
@dopiera, what are the assumptions - I mean: what percentage of verified reads do we expect to deal with - are here?

I don't think we have any named expectations - we should come up with some. I'd guess, the users will use it the following way:

disable altogether
enable for all operations
enable for a negligible fraction of operations (AKA 1%)
enable for a small fraction of operations (AKA 10%)

dopiera

Reviewable status: 3 of 20 files reviewed, 4 unresolved discussions (waiting on @dopiera, @mwalkiewicz, and @prawilny)

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/MirroringConfigurationHelper.java, line 103 at r3 (raw file):

   * <p>Correct values are a integers ranging from 0 to 100 inclusive.
   */
  public static final String MIRRORING_READ_VERIFICATION_RATE =

Could you please make the name of the constant and the name of the configuration field indicate it's a percentage? Hopefully, that makes people less likely to wonder about the proper unit.

mwalkiewicz

Reviewable status: 3 of 20 files reviewed, 4 unresolved discussions (waiting on @dopiera, @mwalkiewicz, and @prawilny)

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/MirroringTable.java, line 122 at r1 (raw file):

Previously, dopiera (Marek Dopiera) wrote…

I'd reorder the last two parameters. My rationale is importance - tracer is only for monitoring but sampling actually affect the table's behavior.

Done.

bigtable-hbase-mirroring-client-1.x-parent/bigtable-hbase-mirroring-client-1.x/src/main/java/com/google/cloud/bigtable/mirroring/hbase1_x/utils/MirroringConfigurationHelper.java, line 103 at r3 (raw file):

Previously, dopiera (Marek Dopiera) wrote…

Could you please make the name of the constant and the name of the configuration field indicate it's a percentage? Hopefully, that makes people less likely to wonder about the proper unit.

Good point. Done.

dopiera

Reviewable status: 3 of 21 files reviewed, 3 unresolved discussions (waiting on @dopiera, @mwalkiewicz, and @prawilny)

* Initial implementation of hbase1.x Mirroring Client (#1) * Initial implementation of hbase1.x Mirroring Client * MirroringTable and AsyncTableWrapper (#4) * MirroringTable and AsyncTableWrapper - Initial implementation of MirroringTable. - Wrapper around HBase 1.x Table enabling scheduling asynchronous operations. - Asynchronous verification of get/exists results. * Add missing licence header to TestResultComparator (#6) * Setup running tests in GitHub Actions (#7) * Asynchronous ResultScanner (#5) * Asynchronous ResultScanner * Implement faillog.Appender (#2) `faillog.Appender` is the lowermost component of the infrastructure for dumping failed mutations in the `MirroringClient`. `faillog/README.md` explains the design decisions in a bit more detail. `faillog.Appender` essentially logs arbitrary data asynchronously, through a separate thread reading from a bounded buffer. * ListenableReferenceCounter (#8) * Count references in asynchronous tasks before closing Table/Scanner. * Flow controller (#10) * Flow Controller * Flow Control strategy based on single requests queue * Mirroring table: writes (#12) * Add missing condition in result comparator (#15) * Add more mirroring config opitons (#16) * MismatchDetector implementation. * FlowControllerStrategy implementation. * Maximal number of outstanding requests used by FlowControllerStrategy. * Primary/Secondary connection implementation option now accepts "default" value which can be used when default HBase Connection implementation should be used by MirroringConnection. * Make AsyncTableWrapper ListenableCloseable (#20) Implements our standard interface for objects that can run callbacks after asynchronous close to simplify reference counting of MirroringTable. * MirroringConnection: use new config options (#17) * MirroringResultScanner improvements (#18) * Count references to MirroringResultScanner Current implementation of MirroringResultScanner doesn't count verificaiton requests that it has scheduled and allows to close instances while verificaiton requests are in-flight. This causes lost verifications. This PR fixes this issue by counting references to MirroringResultScanner instances when scheduling verification requests. Moreover, ListenableClosable interface is implemented for consistency with other classes that use this scheme, because now the MirroringResultScanner instances will be closed asynchronously, when all scheduled requests have finished. * MirroringTable: count references to scanners and secondary wrapper (#21) Current MirroringTable implementaion does not count its references held by MirroringResultScanners and SecondaryAsyncWrapper, thus MirroringConnection consideres it closed before we are sure that all asynchronous operations have completed. This PR adds correct reference counting of MirroringTable based on work done in previously merged PRs. * Result Scanner - ensure verification ordering (#22) Current implementation assumes that next() operations on primary and secondary scanners are called in the same order and uses this assumption to match results for verification. However, next()s on secondary database are called asynchronusly and their order is not defined, which causes invalid mismatch reports. This PR fixes this problem by placing data to be verified - results of next()s called on primary scanner - and details of next() call - number of requested elements - call on a queue. Each asynchronous call to next() is synchonized and pops a single element from that queue. Appropriate next is called based on number of requested elements. Then results of that request and results from the queue are verified. This ensures that results of next()s passed to verification are correctly matched and ordered. * Integration tests (#23) HBase 1.x integration tests * Add trace logging. (#24) * Estimate memory overhead in RequestResourceDescription (#25) * Tests: extract executor service to TestRule (#26) Extract executor service utilities into TestRule to facilitate code reuse in other test classes. * Integration tests: read configuration from xml files * MirroringBufferedMutator * MirroringBufferedMutator: integration tests (#9) * Fix error introduced in rebase (#11) * Obtain resource reservation before scheduling secondary calls (#4) Fixes a bug when secondary database request Future was created before obtaining resources from FlowController. * Integration tests - MirroringTable operations (#10) * MirroringAsyncConfiguration (#5) Add configuration class to be used by MirroringAsyncConnection. * SecondaryWriteErrorConsumer in MirroringTable (#15) Use SecondaryWriteErrorConsumer to handle write errors in secondary database in MirroringTable's writes. * Use Put to implement Increment and Append (#16) * refactor: extract functions using reflection into package utils.reflection * refactor: extract BatchHelpers into utils Extract common part of batch() helpers into a class and add Predicate argument to nested classes' constructors making it possible to reuse the code in 2.x client. * feat: Initial implementation of a subset of asynchronous HBase 2.x Mirroring Client Contains basic implementation of MirroringAsyncConnection and MirroringAsyncTable. * refactor: extract FlowController's request cancellation into a method * fix: Increment ITs fail with Bigtable as primary (#21) We were setting timerange on Increment objects used in integration tests without any reason and Bigtable doesn't support this operation. Setting timerange in ITs was removed. * fix: RequestScheduling should handle rejected resource reservations (#24) Custom FlowControlerStrategy implementations might, contrary to the default implementation, resolve reservation requests with exception, what we should handle by not performing the action that had to acquire the resources. * feat: Add OpenCensus tracing and metrics. (#14) * fix: make BatchHelpers skip verification of empty read results BatchHelpers provides error handling of batch() when there may be some partial results. Before the commit, matching successful reads were redundantly verified if there were none of them. This commit brings back the behaviour from up to 5a29253: when there are no successful matching reads, a MismatchDetector isn't called on empty arrays. * refactor: make MirroringAsync{Connection,Table} use SecondaryWriteErrorConsumerWithMetrics BatchHelpers require using SecondaryWriteErrorConsumerWithMetrics API. * refactor: make AsyncRequestScheduling accept CompletableFuture<ResourceReservation> instead of ResourceReservation This change is split off from commit introducing MirroringAsyncTable#batch() * feat: implement batch() in MirroringAsyncTable Implementation of MirroringAsyncTable's batch() and MirroringAsyncTable's methods such as get(List<Get) and put(List<Put>) using it. * feat: implement failed mutations log (#19) Failed secondary mutations are written to disk in JSON format, which the user can parse programmatically or inspect visually. Each failure is logged as a separate line, which makes it compatible with solutions like logstash. * refactor: split SplitBatchResponse (#40) SplitBatchResponse was refactored into two parts: splitting into reads/writes and failed/successful. This makes the code simpler and easier to maintain. * refactor: extract helper methods in tests (#48) * refactor: remove redundant writeWithControlFlow argument * feat: copy HBase operations' input lists (#57) * refactor: remove redundant field from MirroringConnection (#55) * feat: verification with sampling (#28) * fix: mirror Increment/Append in batch() using Put. (#47) * refactor: Move HBaseOperation into WriteOperationInfo (#68) * refactor: remove redundant parameter from scheduleWriteWithControlFlow (#69) * fix: integration tests - fix build (#70) * fix: count references to batch operations (#63) * fix: close underlying connections when MirroringConnection is closed (#49) * refactor: fix IDE warnings in MirroringAsyncTable test (#64) * fix: integration tests - check if write errors were reported (#71) * feat: make SecondaryWriteErrorConsumer accept error cause and operation (#65) * fix: do not call callbacks with lock held (#53) * refactor: use AccumulatedExceptions where appropriate (#54) * fix: fix key used in verification sampling ITs (#77) * feat: use faillog for handling write errors (#66) * refactor: add utilities for Futures (FutureUtils) * feat: defer closing connections until async operations complete (#37) Mirroring client schedules asynchronous operations - to mirror the mutations and to verify reads. Before this PR, closing the MirroringAsyncConnection would result in closing the underlying connections immediately. This made the pending asynchronous operations fail. This PR defers closing the underlying connections until all pending operations complete. It is achieved by reference counting the operations. * feat: implement AsyncTableBuilder (#42) * feat: implement MirroringAsyncTable#checkAndMutate (#43) * fix: Implement single Append and Increment as Put (#38) * refactor: simplify SecondaryWriteErrorConsumer API (#78) * feat: concurrent writes in MirroringTable (#79) * test: fix failing concurrent write test (#120) * refactor: renames and moves in RequestScheduling (#87) * wip: handover session comments Co-authored-by: Mateusz Walkiewicz <mwalkiewicz@unoperate.com> Co-authored-by: Adam Czajkowski <prawilny@unoperate.com>

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from aad6051 to 166b773 Compare September 29, 2021 10:56

mwalkiewicz requested review from dopiera and prawilny September 30, 2021 07:01

prawilny previously approved these changes Sep 30, 2021

View reviewed changes

dopiera requested changes Sep 30, 2021

View reviewed changes

mwalkiewicz dismissed prawilny’s stale review via 33152f1 October 1, 2021 08:16

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from d68fb2d to e3b982b Compare October 1, 2021 11:28

mwalkiewicz changed the base branch from master to mw/split-batch/1 October 1, 2021 11:33

mwalkiewicz force-pushed the mw/split-batch/1 branch from 08f7dbc to a67f521 Compare October 1, 2021 12:28

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch 5 times, most recently from 185e6b6 to 8acfbab Compare October 2, 2021 16:58

dopiera requested changes Oct 4, 2021

View reviewed changes

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from 8acfbab to 4283053 Compare October 6, 2021 14:08

mwalkiewicz commented Oct 6, 2021

View reviewed changes

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from 4283053 to 48d60cb Compare October 6, 2021 14:09

mwalkiewicz requested a review from dopiera October 8, 2021 09:51

mwalkiewicz changed the base branch from mw/split-batch/1 to master October 8, 2021 09:52

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from 48d60cb to f331fb8 Compare October 11, 2021 05:55

mwalkiewicz requested a review from prawilny October 11, 2021 05:55

feat: verification with sampling

8125561

mwalkiewicz force-pushed the mw/read-verification-sampling/1 branch from f331fb8 to 8125561 Compare October 11, 2021 06:01

dopiera approved these changes Oct 11, 2021

View reviewed changes

mwalkiewicz merged commit 9063744 into master Oct 12, 2021

dopiera pushed a commit that referenced this pull request May 11, 2022

feat: verification with sampling (#28)

103ec7c

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Verification with sampling #28

feat: Verification with sampling #28

Uh oh!

mwalkiewicz commented Sep 29, 2021 •

edited by dopiera

Loading

Uh oh!

prawilny left a comment

Uh oh!

dopiera left a comment

Uh oh!

dopiera left a comment

Uh oh!

mwalkiewicz left a comment

Uh oh!

dopiera left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

feat: Verification with sampling #28

feat: Verification with sampling #28

Uh oh!

Conversation

mwalkiewicz commented Sep 29, 2021 • edited by dopiera Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

prawilny left a comment

Choose a reason for hiding this comment

Uh oh!

dopiera left a comment

Choose a reason for hiding this comment

Uh oh!

dopiera left a comment

Choose a reason for hiding this comment

Uh oh!

mwalkiewicz left a comment

Choose a reason for hiding this comment

Uh oh!

dopiera left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

mwalkiewicz commented Sep 29, 2021 •

edited by dopiera

Loading