Hi,
I was looking at how compaction is invoked from HBaseAdmin (line 1460 in
trunk code)

        final String tableName = tableNameString(tableNameOrRegionName, ct);

        List<Pair<HRegionInfo, ServerName>> pairs
          MetaReader.getTableRegionsAndLocations(ct,

              tableName);

        for (Pair<HRegionInfo, ServerName> pair: pairs) {

          if (pair.getFirst().isOffline()) continue;

          if (pair.getSecond() == null) continue;

          try {

            compact(pair.getSecond(), pair.getFirst(), major, columnFamily);

We issue one compaction request per region.

Looks like we can group regions which reside on the same server and make
one network trip per region server.
Your comment on this would be appreciated.

Since compactions are long running, async jobs anyways, I don't see much
gain in grouping these together. Sure, we might save a round-trip, but
these are neither high-volume calls nor bottle-necked in the RTT of the
request.

It could be nice to see this unified though into a general multi-region
Callable that can do do this same logic cleanly; certainly could think of a
couple places we might want to do that.
-------------------
Jesse Yates
@jesse_yates
jyates.github.com
On Wed, Mar 20, 2013 at 2:14 PM, Ted Yu <[EMAIL PROTECTED]> wrote:

> Hi,
> I was looking at how compaction is invoked from HBaseAdmin (line 1460 in
> trunk code)
>
>         final String tableName = tableNameString(tableNameOrRegionName,
> ct);
>
>         List<Pair<HRegionInfo, ServerName>> pairs >
>           MetaReader.getTableRegionsAndLocations(ct,
>
>               tableName);
>
>         for (Pair<HRegionInfo, ServerName> pair: pairs) {
>
>           if (pair.getFirst().isOffline()) continue;
>
>           if (pair.getSecond() == null) continue;
>
>           try {
>
>             compact(pair.getSecond(), pair.getFirst(), major,
> columnFamily);
>
> We issue one compaction request per region.
>
> Looks like we can group regions which reside on the same server and make
> one network trip per region server.
>
>
> Your comment on this would be appreciated.
>