NAME

Search::Elasticsearch::Client::1_0::Async::Scroll - A helper module for scrolled searches

VERSION

version 5.02

SYNOPSIS

use Search::Elasticsearch::Async;

my $es = Search::Elasticsearch::Async->new;

my $scroll = $es->scroll_helper
    index       => 'my_index',
    search_type => 'scan',
    size        => 500,
    on_start    => \&on_start,
    on_result   => \&on_result,
  | on_results  => \&on_results,
    on_error    => \&on_error
);

$scroll->start->then( sub {say "Done"}, sub { warn @_ } );

sub on_start {
    my $scroll = shift;
    say "Total hits: ". $scroll->total;
}

sub on_result {
    my $doc = shift;
    do_something($doc);
}

sub on_results {
    for my $doc (@_) {
        do_something($doc)
    }
}

sub on_error {
    my $error = shift;
    warn "$error";
}

DESCRIPTION

A scrolled search is a search that allows you to keep pulling results until there are no more matching results, much like a cursor in an SQL database.

Unlike paginating through results (with the from parameter in search()), scrolled searches take a snapshot of the current state of the index. Even if you keep adding new documents to the index or updating existing documents, a scrolled search will only see the index as it was when the search began.

This module is a helper utility that wraps the functionality of the search() and scroll() methods to make them easier to use.

IMPORTANT: Deep scrolling can be expensive. See "DEEP SCROLLING" for more.

This class does Search::Elasticsearch::Client::1_0::Role::Scroll and Search::Elasticsearch::Role::Is_Async.

USE CASES

There are two primary use cases:

Pulling enough results

Perhaps you want to group your results by some field, and you don't know exactly how many results you will need in order to return 10 grouped results. With a scrolled search you can keep pulling more results until you have enough. For instance, you can search emails in a mailing list, and return results grouped by thread_id:

use Promises qw(deferred);

sub find_email_threads {
    my (%groups,@results,$scroll);

    my $d = deferred;

    $scroll = $es->scroll_helper(
        index     => 'my_emails',
        type      => 'email',
        body      => { query => {... some query ... }},
        on_result => sub {
            my $doc = shift;
            my $thread = $doc->{_source}{thread_id};
            unless ($groups{$thread}) {
                $groups{$thread} = [];
                push @results, $groups{$thread};
            }
            push @{$groups{$thread}},$doc;

            # stop collecting if we have 10 results
            if (@results == 10) {
                $scroll->finish;
            }
        }
    );

    $scroll->start->then(
        # resolve with results if completed successfully
        sub { $d->resolve(@results) },

        # reject with error if failed
        sub { $d->reject(@_) }
    );

    return $d->promise;
}

Extracting all documents

Often you will want to extract all (or a subset of) documents in an index. If you want to change your type mappings, you will need to reindex all of your data. Or perhaps you want to move a subset of the data in one index into a new dedicated index. In these cases, you don't care about sort order, you just want to retrieve all documents which match a query, and do something with them. For instance, to retrieve all the docs for a particular client_id:

$es->scroll_helper(
    index       => 'my_index',
    search_type => 'scan',          # important!
    size        => 500,
    body        => {
        query => {
            match => {
                client_id => 123
            }
        }
    },
    on_result => sub { do_something(@_) }
)->start;

Very often the something that you will want to do with these results involves bulk-indexing them into a new index. The easiest way to marry a scrolled search with bulk indexing is to use the "reindex()" in Search::Elasticsearch::Client::1_0::Async::Bulk method.

DEEP SCROLLING

Deep scrolling (and deep pagination) are very expensive in a distributed environment, and the reason they are expensive is that results need to be sorted in a global order.

For example, if we have an index with 5 shards, and we request the first 10 results, each shard has to return its top 10, and then the requesting node (the node that is handling the search request) has to resort these 50 results to return a global top 10. Now, if we request page 1,000 (ie results 10,001 .. 10,010), then each shard has to return 10,010 results, and the requesting node has to sort through 50,050 results just to return 10 of them!

You can see how this can get very heavy very quickly. This is the reason that web search engines never return more than 1,000 results.

Disable sorting for efficient scrolling

The problem with deep scrolling is the sorting phase. If we disable sorting, then we can happily scroll through millions of documents efficiently. The way to do this is to set search_type to scan:

$es->scroll_helper(
    search_type => 'scan',
    size        => 500,
)->start;

Scanning disables sorting and will just return size results from each shard until there are no more results to return. Note: this means that, when querying an index with 5 shards, the scrolled search will pull size * 5 results at a time. If you have large documents or are memory constrained, you will need to take this into account.

METHODS

new()

use Search::Elasticsearch::Async;

my $es = Search::Elasticsearch::Async->new(...);
my $scroll = $es->scroll_helper(
    scroll             => '1m',            # optional
    scroll_in_qs       => 0|1,             # optional

    on_result          => sub {...}        # required
  | on_results         => sub {...}        # required

    on_start           => sub {...}        # optional
    on_error           => sub {...}        # optional
    %search_params,
);
$scroll->start;

The "scroll_helper()" in Search::Elasticsearch::Client::1_0::Direct method loads Search::Elasticsearch::Client::1_0::Async::Scroll class and calls "new()", passing in any arguments.

You can specify a scroll duration (which defaults to "1m"), scroll_in_qs (which defaults to false), and any of the listed callbacks. Any other parameters are passed directly to "search()" in Search::Elasticsearch::Client::1_0::Direct.

The scroll duration tells Elasticearch how long it should keep the scroll alive. Note: this duration doesn't need to be long enough to process all results, just long enough to process a single batch of results. The expiry gets renewed for another scroll period every time new a new batch of results is retrieved from the cluster.

By default, the scroll_id is passed as the body to the scroll request. To send it in the query string instead, set scroll_in_qs to a true value, but be aware: when querying very many indices, the scroll ID can become too long for intervening proxies.

The scroll request uses GET by default. To use POST instead, set send_get_body_as to POST.

Callbacks

You must specify either an on_result callback or an on_results callback.

on_result and on_results

The on_result callback is called once for every result that is received.

sub on_result {
    my $doc = shift;
    do_something($doc);
}

Alternatively, you can specify an on_results callback which is called once for every set of results returned by Elasticsearch:

sub on_results {
    for my $doc (@_) {
        do_something($doc)
    }
}

If either on_result or on_results returns a new Promise, processing of further results will be paused until the promise has been rejected or resolved.

on_start

The on_start callback is called after the first request has completed, at which stage the properties like total(), aggregations(), etc will have been populated.

on_error

The on_error callback is called if any error occurs. The default implementation warns about the error, and rethrows it.

sub on_error { warn "Scroll error: @_"; die @_ }

If you wish to handle (and surpress) certain errors, then don't call die(), eg:

sub on_error {
    my $error = shift;
    if ($error =~/SomeCatchableError/) {
        # do something to handle error
    }
    else {
        # rethrow error
        die $error;
    }
}

start()

$scroll->start
       ->then( \&success, \&failure );

The start() method starts the scroll and returns a Promise which will be resolved when the scroll completes (or "finish()" is called), or rejected if any errors remain unhandled.

finish()

$scroll->finish;

The finish() method clears out the buffer, sets "is_finished()" to true and tries to clear the scroll_id on Elasticsearch. This API is only supported since v0.90.5, but the call to clear_scroll is wrapped in an eval so the finish() method can be safely called with any version of Elasticsearch.

When the $scroll instance goes out of scope, "finish()" is called automatically if required.

is_finished()

$bool = $scroll->is_finished;

A flag which returns true if all results have been processed or "finish()" has been called.

INFO ACCESSORS

The information from the original search is returned via the accessors below. These values can be accessed in the on_start callback:

total

The total number of documents that matched your query.

max_score

The maximum score of any documents in your query.

aggregations

Any aggregations that were specified, or undef

facets

Any facets that were specified, or undef

suggest

Any suggestions that were specified, or undef

took

How long the original search took, in milliseconds

took_total

How long the original search plus all subsequent batches took, in milliseconds. This value can only be checked once the scroll has completed.

SEE ALSO

AUTHOR

Clinton Gormley <drtech@cpan.org>

COPYRIGHT AND LICENSE

This software is Copyright (c) 2017 by Elasticsearch BV.

This is free software, licensed under:

The Apache License, Version 2.0, January 2004