Using Hash Clusters

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Cluster Blocks
Table Directory Row Directory
0
1
0
1
2
3
4

• Table directory 0
1
2

• partitions the row directory

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Cluster Key Changes
• Performed by recursive/service transactions
• commit independent of parent/user transaction
• improves concurrency
• Single threaded
• to prevent incompatible key changes
• always use ITL slot 1
• slot 1 is reserved - user transactions never use it
• wait in shared mode for TX lock if slot is busy

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Index Cluster Inserts
• Key can be inserted to block if
• number of keys < (data area size - 14) / SIZE
• available space > SIZE - 27
• key and data row pieces fit below PCTFREE
• Row inserts
• lookup key in cluster index
• insert to (last) data block if space permits
• try up to 3 previous blocks in the key chain
• link a freelist block into the key chain & insert
Unix + Oracle =
Steve Adams
Ixora [email protected]
 Index Cluster Deletes
• Data row piece headers have a key index
• Keys have current & pending row counts
• both incremented on insert
• pending decremented on delete
• current decremented on stub cleanout
• Flushable keys
• current count was, or pending count is zero
• when inserting a new key, try to clean out all
committed stubs and flush keys if possible
Unix + Oracle =
Steve Adams
Ixora [email protected]
 Demonstration
• Flushable keys
• Key insert restrictions

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Demonstration
• Index clusters row inserts

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Cluster Key Scans
• Purpose
• to fetch all rows for a clustered table
with a known cluster key value

• Buffer gets
• visit all blocks in the key chain (reverse order)
• scan row directory entries for the table
and check cluster key index for all rows

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Assessing Cluster Key Chaining
• Detection
• compare cluster key scan block gets
to the cluster key scans statistic
• maybe DBA_CLUSTERS .
AVG_BLOCKS_PER_KEY
• Fix
• rename existing tables and create new cluster
• use PL/SQL iterate over the cluster keys
and copy all rows for each key in turn
Unix + Oracle =
Steve Adams
Ixora [email protected]
 Hash Theory
• TBD

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Hash Clusters
• Fixed blocks
• formatted during cluster creation
• hash key value implied by location
• keys never flushed; blocks never freed
• no data row counts, or previous key rowid needed
• Chained blocks
• just like index cluster blocks
• key data is hash key value

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Demonstration
• Hash cluster

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Single Table Hash Clusters
• Better “single row per hash key”
• Reserved slots
• one row slot per key
• cluster key scans only inspect that slot
iff no keys have more than one row
• otherwise scan all row slots for table
• “A” flag in data layer flag byte
• no more than one row per key

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Demonstration
• Single table hash cluster

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Hash Functions
• Internal hash function
• returns 31-bit positive integer
• good distribution, but not perfect
• mod(hash value, HASHKEYS)
• HASH IS column_name
• must be integer column
• saves hash computation
• no false collisions

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Demonstration
• Single table hash cluster without HASH IS

Unix + Oracle =
Steve Adams
Ixora [email protected]
 Hash Cluster Data Density
• Use SINGLE TABLE syntax if possible
• Otherwise
• 2K tablespace block size
• no SIZE parameter
• HASHKEYS is number of fixed blocks needed
• primary key columns first in row
• Benefits
• minimize CPU cost of cluster key scans
• maximize data density - only one key per block
Unix + Oracle =
Steve Adams
Ixora [email protected]