accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Slater, David M." <>
Subject RE: Improving Batchscanner Performance
Date Tue, 20 May 2014 18:24:38 GMT

The data is not significantly larger than the rows that I'm fetching. in terms of bandwidth,
the data returned is at least 2 orders of magnitude smaller than the ingest rate, so I don't
think it's a network issue.

I'm guessing, as Bob suggested, that it has to do with fetching a "random" set of rows each
time. I had assumed that the batchscanner would take the Collection of ranges (when setting
batchScanner.setRanges()), sort them, and then fetch data based on tablet splits. I'm guessing,
based on the discussion, that it is not done that way. 

Does the BatchScanner fetch rows based on the ordering of the Collection?


-----Original Message-----
From: Josh Elser [] 
Sent: Tuesday, May 20, 2014 1:59 PM
Subject: Re: Improving Batchscanner Performance

You actually stated it exactly here:

 > I complete the first scan in its entirety

Loading the data into a Collection also implies that you're loading the complete set of rows
and blocking until you find all rows, or until you fetch all of the data.

 > Collection<Text> rows = getRowIDs(new Range("minRow", "maxRow"), new Text("index"),
"mytable", 10, 10000);  > Collection<byte[]> data = getRowData(rows, "mytable", 10);

Both the BatchScanner and Scanner are returning KeyValue pairs in "batches". The client talks
to server(s), reads some data and returns it to you. By virtue of you loading these results
from the Iterator into a Collection, you are consuming *all* results before proceeding to
fetch the data for the rows.

Now, if, like you said, looking up the rows is drastically faster than fetching the data,
there's a question as to why this is. Is it safe to assume that the data is much larger than
the rows you're fetching? Have you tried to see what the throughput of fetching this data
is? If it's bounded by network speed, you could try compressing the data in an iterator server-side
before returning it to the client.

You could also consider the locality of the rows that you're fetching -- are you fetching
a "random" set of rows each time and paying a penalty of talking to each server to fetch the
data when you could ammortize the cost if you fetched the data for rows that are close together.
A large amount of data being returned is likely going to trump the additional cost in talking
to many servers.

On 5/20/14, 1:51 PM, Slater, David M. wrote:
> Hi Josh,
> I should have clarified - I am using a batchscanner for both lookups. I had thought of
putting it into two different threads, but the first scan is typically an order of magnitude
faster than the second.
> The logic for upperbounding the results returned is outside of the method I provided.
Since there is a one-to-one relationship between rowIDs and records on the second scan, I
just limit the number of rows I send to this method.
> As for blocking, I'm not sure exactly what you mean. I complete the first scan in its
entirety, which  before entering this method with the collection of Text rowIDs. The method
for that is:
> public Collection<Text> getRowIDs(Collection<Range> ranges, Text term, String
tablename, int queryThreads, int limit) throws TableNotFoundException {
>          Set<Text> guids = new HashSet<Text>();
>          if (!ranges.isEmpty()) {
>              BatchScanner scanner = conn.createBatchScanner(tablename, new Authorizations(),
>              scanner.setRanges(ranges);
>              scanner.fetchColumnFamily(term);
>              for (Map.Entry<Key, Value> entry : scanner) {
>                  guids.add(entry.getKey().getColumnQualifier());
>                  if (guids.size() > limit) {
>                      return null;
>                  }
>              }
>              scanner.close();
>          }
>          return guids;
>      }
> Essentially, my query does:
> Collection<Text> rows = getRowIDs(new Range("minRow", "maxRow"), new 
> Text("index"), "mytable", 10, 10000); Collection<byte[]> data = 
> getRowData(rows, "mytable", 10);
> -----Original Message-----
> From: Josh Elser []
> Sent: Tuesday, May 20, 2014 1:32 PM
> To:
> Subject: Re: Improving Batchscanner Performance
> Hi David,
> Absolutely. What you have here is a classic producer-consumer model.
> Your BatchScanner is producing results, which you then consume by your scanner, and ultimately
return those results to the client.
> The problem with your below implementation is that you're not going to be polling your
batchscanner as aggressively as you could be. You are blocking while you can fetch each of
those new Ranges from the Scanner before fetching new ranges. Have you considered splitting
up the BatchScanner and Scanner code into two different threads?
> You could easily use a ArrayBlockingQueue (or similar) to pass results from the BatchScanner
to the Scanner. I would imagine that this would give you a fair improvement in performance.
> Also, it doesn't appear that there's a reason you can't use a BatchScanner for both lookups?
> One final warning, your current implementation could also hog heap very badly if your
batchscanner returns too many records. The producer/consumer I proposed should help here a
little bit, but you should still be asserting upper-bounds to avoid running out of heap space
in your client.
> On 5/20/14, 1:10 PM, Slater, David M. wrote:
>> Hey everyone,
>> I'm trying to improve the query performance of batchscans on my data table. I first
scan over index tables, which returns a set of rowIDs that correspond to the records I am
interested in. This set of records is fairly randomly (and uniformly) distributed across a
large number of tablets, due to the randomness of the UID and the query itself. Then I want
to scan over my data table, which is setup as follows:
>> row     		colFam      	colQual     	value
>> rowUID  	 --          		--          		byte[] of data
>> These records are fairly small (100s of bytes), but numerous (I may return 50000
or more). The method I use to obtain this follows. Essentially, I turn the rows returned from
the first query into a set of ranges to input into the batchscanner, and then return those
rows, retrieving the value from them.
>> // returns the data associated with the given collection of rows
>>       public Collection<byte[]> getRowData(Collection<Text> rows, Text
dataType, String tablename, int queryThreads) throws TableNotFoundException {
>>           List<byte[]> values = new ArrayList<byte[]>(rows.size());
>>           if (!rows.isEmpty()) {
>>               BatchScanner scanner = conn.createBatchScanner(tablename, new Authorizations(),
>>               List<Range> ranges = new ArrayList<Range>();
>>               for (Text row : rows) {
>>                   ranges.add(new Range(row));
>>               }
>>               scanner.setRanges(ranges);
>>               for (Map.Entry<Key, Value> entry : scanner) {
>>                   values.add(entry.getValue().get());
>>               }
>>               scanner.close();
>>           }
>>           return values;
>>       }
>> Is there a more efficient way to do this? I have index caches and bloom filters enabled
(data caches are not), but I still seem to have a long query lag. Any thoughts on how I can
improve this?
>> Thanks,
>> David

View raw message