Andrew Purtell created HBASE-10742:
--------------------------------------
Summary: Data temperature aware compaction policy
Key: HBASE-10742
URL: https://issues.apache.org/jira/browse/HBASE-10742
Project: HBase
Issue Type: Brainstorming
Reporter: Andrew Purtell
Reading "Identifying Hot and Cold Data in Main-Memory Databases" (Levandoski,
Larson, and Stoica), it occurred to me that some of the motivation applies to
HBase and some of the results can inform a data temperature aware compaction
policy implementation.
We also wish to optimize retention of cells in the working set in memory, in
blockcache.
We can also consider further and related performance optimizations in HBase
that awareness of hot and cold data can enable, even for cases where the
working set does not fit in memory. If we could partition HFiles into hot and
cold (cold+lukewarm) and move cells between them at compaction time, then we
could:
- Migrate hot HFiles onto alternate storage tiers with improved read latency
and throughput characteristics. This has been discussed before on HBASE-6572.
Or, migrate cold HFiles to an archival tier.
- Preload hot HFiles into blockcache to increase cache hit rates, especially
when regions are first brought online. And/or add another LRU priority to
increase the likelihood of retention of blocks in hot HFiles. This could be
sufficiently different from ARC to avoid issues there.
- Reduce the compaction priorities of cold HFiles, with proportional reduction
in priority IO and write amplification, since cold files would less frequently
participate in reads.
Levandoski et. al. describe determining data temperature with low overhead
using an out of band estimation process running in the background over an
access log. We could consider logging reads along with mutations and similarly
process the result in the background. The WAL could be overloaded to carry
access log records, or we could follow the approach described in the paper and
maintain an in memory access log only.
{quote}
We chose the offline approach for several reasons. First, as mentioned earlier,
the overhead of even the simplest caching scheme is very high. Second, the
offline approach is generic and requires minimum changes to the database
engine. Third, logging imposes very little overhead during normal operation.
Finally, it allows flexibility in when, where, and how to analyze the log and
estimate access frequencies. For instance, the analysis can be done on a
separate machine, thus reducing overhead on the system running the
transactional workloads.
{quote}
Importantly, they only log a sample of all accesses.
{quote}
To implement sampling, we have each worker thread flip a biased coin before
starting a new query (where bias correlates with sample rate). The thread
records its accesses in log buffers (or not) based on the outcome of the coin
flip. In Section V, we report experimental results showing that sampling 10% of
the accesses reduces the accuracy by only 2.5%,
{quote}
Likewise we would only record a subset of all accesses to limit overheads.
The offline process estimates access frequencies over discrete time slices
using exponential smoothing. (Markers representing time slice boundaries are
interleaved with access records in the log.) Forward and backward
classification algorithms are presented. The forward algorithm requires a full
scan over the log and storage proportional to the number of unique cell
addresses, while the backward algorithm requires reading a least the tail of
the log in reverse order.
If we overload the WAL to carry the access log, offline data temperature
estimation can piggyback as a WAL listener. If replication is also enabled then
we might even consolidate both forward scans through the WAL into a single
sequential read. The forward algorithm would then be a natural choice. The
HBase master is fairly idle most of the time and less memory hungry as a
regionserver, at least in today's architecture. We could probably get away with
considering only row+family as a unique coordinate to minimize space overhead.
Or if instead we maintain the access logs in memory at the RegionServer, then
there is a parallel formulation and we could benefit from the reverse
algorithm's ability to terminate early once confidence bounds are reached and
backwards scanning IO wouldn't be a concern. This handwaves over a lot of
details.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
