6.3. Rowkey Design

6.3.1. Hotspotting

Rows in HBase are sorted lexicographically by row key. This design optimizes for scans, allowing you to store related rows, or rows that will be read together, near each other. However, poorly designed row keys are a common source of hotspotting. Hotspotting occurs when a large amount of client traffic is directed at one node, or only a few nodes, of a cluster. This traffic may represent reads, writes, or other operations. The traffic overwhelms the single machine responsible for hosting that region, causing performance degradation and potentially leading to region unavailability. This can also have adverse effects on other regions hosted by the same region server as that host is unable to service the requested load. It is important to design data access patterns such that the cluster is fully and evenly utilized.

To prevent hotspotting on writes, design your row keys such that rows that truly do need to be in the same region are, but in the bigger picture, data is being written to multiple regions across the cluster, rather than one at a time. Some common techniques for avoiding hotspotting are described below, along with some of their advantages and drawbacks.

Salting. Salting in this sense has nothing to do with cryptography, but refers to adding random data to the start of a row key. In this case, salting refers to adding a randomly-assigned prefix to the row key to cause it to sort differently than it otherwise would. The number of possible prefixes correspond to the number of regions you want to spread the data across. Salting can be helpful if you have a few "hot" row key patterns which come up over and over amongst other more evenly-distributed rows. Consider the following example, which shows that salting can spread write load across multiple regionservers, and illustrates some of the negative implications for reads.

Example 6.1. Salting Example

Suppose you have the following list of row keys, and your table is split such that there is one region for each letter of the alphabet. Prefix 'a' is one region, prefix 'b' is another. In this table, all rows starting with 'f' are in the same region. This example focuses on rows with keys like the following:

foo0001
foo0002
foo0003
foo0004          
        

Now, imagine that you would like to spread these across four different regions. You decide to use four different salts: a, b, c, and d. In this scenario, each of these letter prefixes will be on a different region. After applying the salts, you have the following rowkeys instead. Since you can now write to four separate regions, you theoretically have four times the throughput when writing that you would have if all the writes were going to the same region.

a-foo0003
b-foo0001
c-foo0004
d-foo0002          
        

Then, if you add another row, it will randomly be assigned one of the four possible salt values and end up near one of the existing rows.

a-foo0003
b-foo0001
c-foo0003
c-foo0004
d-foo0002        
        

Since this assignment will be random, you will need to do more work if you want to retrieve the rows in lexicographic order. In this way, salting attempts to increase throughput on writes, but has a cost during reads.


Hashing. Instead of a random assignment, you could use a one-way hash that would cause a given row to always be "salted" with the same prefix, in a way that would spread the load across the regionservers, but allow for predictability during reads. Using a deterministic hash allows the client to reconstruct the complete rowkey and use a Get operation to retrieve that row as normal.

Example 6.2. Hashing Example

Given the same situation in the salting example above, you could instead apply a one-way hash that would cause the row with key foo0003 to always, and predictably, receive the a prefix. Then, to retrieve that row, you would already know the key. You could also optimize things so that certain pairs of keys were always in the same region, for instance.


Reversing the Key. A third common trick for preventing hotspotting is to reverse a fixed-width or numeric row key so that the part that changes the most often (the least significant digit) is first. This effectively randomizes row keys, but sacrifices row ordering properties.

See https://communities.intel.com/community/itpeernetwork/datastack/blog/2013/11/10/discussion-on-designing-hbase-tables, and article on Salted Tables from the Phoenix project, and the discussion in the comments of HBASE-11682 for more information about avoiding hotspotting.

6.3.2.  Monotonically Increasing Row Keys/Timeseries Data

In the HBase chapter of Tom White's book Hadoop: The Definitive Guide (O'Reilly) there is a an optimization note on watching out for a phenomenon where an import process walks in lock-step with all clients in concert pounding one of the table's regions (and thus, a single node), then moving onto the next region, etc. With monotonically increasing row-keys (i.e., using a timestamp), this will happen. See this comic by IKai Lan on why monotonically increasing row keys are problematic in BigTable-like datastores: monotonically increasing values are bad. The pile-up on a single region brought on by monotonically increasing keys can be mitigated by randomizing the input records to not be in sorted order, but in general it's best to avoid using a timestamp or a sequence (e.g. 1, 2, 3) as the row-key.

If you do need to upload time series data into HBase, you should study OpenTSDB as a successful example. It has a page describing the schema it uses in HBase. The key format in OpenTSDB is effectively [metric_type][event_timestamp], which would appear at first glance to contradict the previous advice about not using a timestamp as the key. However, the difference is that the timestamp is not in the lead position of the key, and the design assumption is that there are dozens or hundreds (or more) of different metric types. Thus, even with a continual stream of input data with a mix of metric types, the Puts are distributed across various points of regions in the table.

See Section 6.11, “Schema Design Case Studies” for some rowkey design examples.

6.3.3. Try to minimize row and column sizes

Or why are my StoreFile indices large?

In HBase, values are always freighted with their coordinates; as a cell value passes through the system, it'll be accompanied by its row, column name, and timestamp - always. If your rows and column names are large, especially compared to the size of the cell value, then you may run up against some interesting scenarios. One such is the case described by Marc Limotte at the tail of HBASE-3551 (recommended!). Therein, the indices that are kept on HBase storefiles (Section 9.7.6.4, “StoreFile (HFile)”) to facilitate random access may end up occupyng large chunks of the HBase allotted RAM because the cell value coordinates are large. Mark in the above cited comment suggests upping the block size so entries in the store file index happen at a larger interval or modify the table schema so it makes for smaller rows and column names. Compression will also make for larger indices. See the thread a question storefileIndexSize up on the user mailing list.

Most of the time small inefficiencies don't matter all that much. Unfortunately, this is a case where they do. Whatever patterns are selected for ColumnFamilies, attributes, and rowkeys they could be repeated several billion times in your data.

See Section 9.7.6.6, “KeyValue” for more information on HBase stores data internally to see why this is important.

6.3.3.1. Column Families

Try to keep the ColumnFamily names as small as possible, preferably one character (e.g. "d" for data/default).

See Section 9.7.6.6, “KeyValue” for more information on HBase stores data internally to see why this is important.

6.3.3.2. Attributes

Although verbose attribute names (e.g., "myVeryImportantAttribute") are easier to read, prefer shorter attribute names (e.g., "via") to store in HBase.

See Section 9.7.6.6, “KeyValue” for more information on HBase stores data internally to see why this is important.

6.3.3.3. Rowkey Length

Keep them as short as is reasonable such that they can still be useful for required data access (e.g., Get vs. Scan). A short key that is useless for data access is not better than a longer key with better get/scan properties. Expect tradeoffs when designing rowkeys.

6.3.3.4. Byte Patterns

A long is 8 bytes. You can store an unsigned number up to 18,446,744,073,709,551,615 in those eight bytes. If you stored this number as a String -- presuming a byte per character -- you need nearly 3x the bytes.

Not convinced? Below is some sample code that you can run on your own.

// long
//
long l = 1234567890L;
byte[] lb = Bytes.toBytes(l);
System.out.println("long bytes length: " + lb.length);   // returns 8

String s = "" + l;
byte[] sb = Bytes.toBytes(s);
System.out.println("long as string length: " + sb.length);    // returns 10

// hash
//
MessageDigest md = MessageDigest.getInstance("MD5");
byte[] digest = md.digest(Bytes.toBytes(s));
System.out.println("md5 digest bytes length: " + digest.length);    // returns 16

String sDigest = new String(digest);
byte[] sbDigest = Bytes.toBytes(sDigest);
System.out.println("md5 digest as string length: " + sbDigest.length);    // returns 26
        

Unfortunately, using a binary representation of a type will make your data harder to read outside of your code. For example, this is what you will see in the shell when you increment a value:

hbase(main):001:0> incr 't', 'r', 'f:q', 1
COUNTER VALUE = 1

hbase(main):002:0> get 't', 'r'
COLUMN                                        CELL
 f:q                                          timestamp=1369163040570, value=\x00\x00\x00\x00\x00\x00\x00\x01
1 row(s) in 0.0310 seconds
        

The shell makes a best effort to print a string, and it this case it decided to just print the hex. The same will happen to your row keys inside the region names. It can be okay if you know what's being stored, but it might also be unreadable if arbitrary data can be put in the same cells. This is the main trade-off.

6.3.4. Reverse Timestamps

Reverse Scan API

HBASE-4811 implements an API to scan a table or a range within a table in reverse, reducing the need to optimize your schema for forward or reverse scanning. This feature is available in HBase 0.98 and later. See https://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/Scan.html#setReversed%28boolean for more information.

A common problem in database processing is quickly finding the most recent version of a value. A technique using reverse timestamps as a part of the key can help greatly with a special case of this problem. Also found in the HBase chapter of Tom White's book Hadoop: The Definitive Guide (O'Reilly), the technique involves appending (Long.MAX_VALUE - timestamp) to the end of any key, e.g., [key][reverse_timestamp].

The most recent value for [key] in a table can be found by performing a Scan for [key] and obtaining the first record. Since HBase keys are in sorted order, this key sorts before any older row-keys for [key] and thus is first.

This technique would be used instead of using Section 6.4, “ Number of Versions ” where the intent is to hold onto all versions "forever" (or a very long time) and at the same time quickly obtain access to any other version by using the same Scan technique.

6.3.5. Rowkeys and ColumnFamilies

Rowkeys are scoped to ColumnFamilies. Thus, the same rowkey could exist in each ColumnFamily that exists in a table without collision.

6.3.6. Immutability of Rowkeys

Rowkeys cannot be changed. The only way they can be "changed" in a table is if the row is deleted and then re-inserted. This is a fairly common question on the HBase dist-list so it pays to get the rowkeys right the first time (and/or before you've inserted a lot of data).

6.3.7. Relationship Between RowKeys and Region Splits

If you pre-split your table, it is critical to understand how your rowkey will be distributed across the region boundaries. As an example of why this is important, consider the example of using displayable hex characters as the lead position of the key (e.g., "0000000000000000" to "ffffffffffffffff"). Running those key ranges through Bytes.split (which is the split strategy used when creating regions in HBaseAdmin.createTable(byte[] startKey, byte[] endKey, numRegions) for 10 regions will generate the following splits...

48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48                                // 0
54 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10                 // 6
61 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -68                 // =
68 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -126  // D
75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 72                                // K
82 18 18 18 18 18 18 18 18 18 18 18 18 18 18 14                                // R
88 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -44                 // X
95 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -102                // _
102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102                // f
      

... (note: the lead byte is listed to the right as a comment.) Given that the first split is a '0' and the last split is an 'f', everything is great, right? Not so fast.

The problem is that all the data is going to pile up in the first 2 regions and the last region thus creating a "lumpy" (and possibly "hot") region problem. To understand why, refer to an ASCII Table. '0' is byte 48, and 'f' is byte 102, but there is a huge gap in byte values (bytes 58 to 96) that will never appear in this keyspace because the only values are [0-9] and [a-f]. Thus, the middle regions regions will never be used. To make pre-spliting work with this example keyspace, a custom definition of splits (i.e., and not relying on the built-in split method) is required.

Lesson #1: Pre-splitting tables is generally a best practice, but you need to pre-split them in such a way that all the regions are accessible in the keyspace. While this example demonstrated the problem with a hex-key keyspace, the same problem can happen with any keyspace. Know your data.

Lesson #2: While generally not advisable, using hex-keys (and more generally, displayable data) can still work with pre-split tables as long as all the created regions are accessible in the keyspace.

To conclude this example, the following is an example of how appropriate splits can be pre-created for hex-keys:.

public static boolean createTable(HBaseAdmin admin, HTableDescriptor table, byte[][] splits)
throws IOException {
  try {
    admin.createTable( table, splits );
    return true;
  } catch (TableExistsException e) {
    logger.info("table " + table.getNameAsString() + " already exists");
    // the table already exists...
    return false;
  }
}

public static byte[][] getHexSplits(String startKey, String endKey, int numRegions) {
  byte[][] splits = new byte[numRegions-1][];
  BigInteger lowestKey = new BigInteger(startKey, 16);
  BigInteger highestKey = new BigInteger(endKey, 16);
  BigInteger range = highestKey.subtract(lowestKey);
  BigInteger regionIncrement = range.divide(BigInteger.valueOf(numRegions));
  lowestKey = lowestKey.add(regionIncrement);
  for(int i=0; i < numRegions-1;i++) {
    BigInteger key = lowestKey.add(regionIncrement.multiply(BigInteger.valueOf(i)));
    byte[] b = String.format("%016x", key).getBytes();
    splits[i] = b;
  }
  return splits;
}
comments powered by Disqus