11. Tuning python-oracledb

Some general tuning tips are:

11.1. Tuning Fetch Performance

To improve application performance and scalability, you can set Cursor.prefetchrows and Cursor.arraysize to adjust python-oracledb’s internal buffers used for SELECT and REF CURSOR query results. Increasing the sizes causes bigger batches of rows to be fetched by each internal request to the database. This reduces the number of round-trips required. The performance benefit of larger buffer sizes can be significant when fetching lots of rows, or when using a slow network. The database also benefits from reduced overheads.

Internally, python-oracledb always does row prefetching and array fetching. Row prefetching is when rows are returned from the database in the round-trip used for initial statement execution. Rows are then available in an internal buffer for the application to consume. Without prefetching, the first set of rows would have to fetched from the database using a separate round-trip. When all rows in the prefetch buffer have been consumed by the application, the next set of rows are internally fetched into a buffer using an array fetch which takes one round-trip. These rows are made available to the application when it wants them. After the initial array fetch buffer has been emptied, further array fetches into the buffer occur as needed to completely retrieve all rows from the database.

Changing the buffer sizes with prefetchrows and arraysize affects the internal buffering behavior of Cursor.fetchone(), Cursor.fetchmany(), and Cursor.fetchall(). The attribute values do not affect how rows are returned to the application with the exception of Cursor.fetchmany(), where Cursor.arraysize is also used as the default for its size parameter.

An overhead of row prefetching in python-oracledb Thick mode is the use of a separate buffer from the array fetch buffer, increasing the memory requirements. Also, an additional data copy from Oracle Client’s prefetch buffer is performed. These costs may outweigh the benefits of using prefetching in some cases.

SELECT queries that return LOB objects and similar types will never prefetch rows, so the prefetchrows value is ignored in those cases.

The attributes do not affect data insertion. To reduce round-trips for DML statements, see Batch Statement and Bulk Copy Operations.

11.1.1. Choosing values for arraysize and prefetchrows

Tune “array fetching” with Cursor.arraysize. Tune “row prefetching” with Cursor.prefetchrows. Set these before calling Cursor.execute(). Also, see Application Default Prefetchrows and Arraysize Values.

The best arraysize and prefetchrows values can be found by benchmarking your application under production load. Here are starting suggestions for four common scenarios:

  • Scenario 1: To tune queries that return an unknown and large number of rows, increase the arraysize value from its default of 100 until you are satisfied with performance, memory, and round-trip usage. For example:

    cursor = connection.cursor()

# cursor.prefetchrows = 2  # generally leave this at its default
cursor.arraysize = 1000

for row in cursor.execute("select * from very_big_table"):
    print(row)

    In general for this scenario, leave prefetchrows at its default value. If you do change it, then set Cursor.arraysize to a bigger value. Do not make the sizes unnecessarily large.

  • Scenario 2: If you are fetching a fixed number of rows, set arraysize to the number of expected rows, and set prefetchrows to one greater than this value. Adding one removes the need for a round-trip to check for end-of-fetch. For example, if you are using a SELECT query to retrieve 20 rows, perhaps to display a page of data, then set prefetchrows to 21 and arraysize to 20:

    cursor = connection.cursor()

cursor.prefetchrows = 21
cursor.arraysize = 20

for row in cursor.execute("""
    select last_name
       from employees
       order by last_name
       offset 0 rows fetch next 20 rows only"""):
    print(row)

    This will return all rows for the query in one round-trip.

  • Scenario 3: If the number of rows returned by a SELECT statement varies from execution to execution in your application but is never large, you can use a similar strategy to Scenario 2.

    Choose arraysize and prefetchrows values that work well for the expected maximum number of rows.

  • Scenario 4: If you know that a SELECT query returns just one row then set arraysize to 1 to minimize memory usage. The default prefetch value of 2 allows minimal round-trips for single-row queries:

    cursor = connection.cursor()

cursor.arraysize = 1

cursor.execute("select * from MyTable where id = 1"):
row = cursor.fetchone()
print(row)

Round-trips Required for Fetching Rows

The following table shows the number of round-trips required to fetch various numbers of rows with different prefetchrows and arraysize values.

Table 11.1 Effect of prefetchrows and arraysize on the number of round-trips

Number of rows

prefetchrows

arraysize

Round-trips

1

2

100

1

100

2

100

2

1000

2

100

11

10000

2

100

101

10000

2

1000

11

10000

1000

1000

11

20

20

20

2

20

21

20

1

The number of round-trips will be the same regardless of which python-oracledb method is used to fetch query results.

11.1.1.1. Application Default Prefetchrows and Arraysize Values

Application-wide defaults can be set using oracledb.defaults.prefetchrows and oracledb.defaults.arraysize, for example:

import oracledb

oracledb.defaults.prefetchrows = 1000
oracledb.defaults.arraysize    = 1000

When using python-oracledb in Thick mode, prefetching can also be tuned in an external oraaccess.xml file, which may be useful for tuning an application when modifying its code is not feasible.

Setting the sizes with oracledb.defaults attributes or with oraaccess.xml will affect the whole application, so it should not be the first tuning choice.

11.1.1.2. Changing Prefetchrows and Arraysize for Re-executed Statements

In python-oracledb, the prefetchrows and arraysize values are only examined when a statement is executed the first time. To change the values for a re-executed statement, create a new cursor. For example, to change arraysize:

array_sizes = (10, 100, 1000)
for size in array_sizes:
    cursor = connection.cursor()
    cursor.arraysize = size
    start = time.time()
    cursor.execute(sql).fetchall()
    elapsed = time.time() - start
    print("Time for", size, elapsed, "seconds")

11.1.1.3. Avoiding Premature Prefetching

There are two cases that will benefit from setting prefetchrows to zero:

  • When passing a python-oracledb cursor into PL/SQL. Setting prefetchrows to 0 can stop rows being prematurely (and silently) fetched into the python-oracledb internal buffer, making those rows unavailable to the PL/SQL REF CURSOR parameter:

    refcursor = connection.cursor()
refcursor.prefetchrows = 0
refcursor.execute("select ...")
cursor.callproc("myproc", [refcursor])

  • When querying a PL/SQL function that uses PIPE ROW to emit rows at intermittent intervals. By default, several rows needs to be emitted by the function before python-oracledb can return them to the application. Setting prefetchrows to 0 helps give a consistent flow of data to the application.

11.1.2. Tuning Fetching from REF CURSORS

The internal buffering and performance of fetching data from REF CURSORS can be tuned by setting the value of arraysize before rows are fetched from the cursor. The prefetchrows value is ignored when fetching from REF CURSORS.

For example:

ref_cursor = connection.cursor()
cursor.callproc("myrefcursorproc", [ref_cursor])

ref_cursor.arraysize = 1000
print("Sum of IntCol for", num_rows, "rows:")
for row in ref_cursor:
    sum_rows += row[0]
print(sum_rows)

The Cursor.arraysize` value can also be set before calling the procedure:

ref_cursor = connection.cursor()
ref_cursor.arraysize = 1000

cursor.callproc("myrefcursorproc", [ref_cursor])
for row in ref_cursor:
    . . .

Also see Avoiding Premature Prefetching.

11.1.3. Tuning Fetching for Data Frames

When fetching data frames with Connection.fetch_df_all() or Connection.fetch_df_batches(), tuning of data transfer across the network is controlled by the respective methods arraysize or size parameters.

Any oracledb.defaults.prefetchrows value is ignored since these methods always set the internal prefetch size to the relevant arraysize or size value.

11.1.4. Parallelizing Data Fetches from a Single Table

Before trying to improve the performance of querying a single table by issuing multiple SELECT queries in multiple threads, where each query extracts a different range of data, you should do careful benchmarking.

Factors that will impact such a solution:

  • How heavily loaded is the database? The parallel solution may appear to be fast but it could be inefficient, thereby impacting, or eventually being limited by, everyone else.

  • A single python-oracledb connection can only do one database operation at a time, so you need to use a different connection for each executed SQL statement, for example, using connections from a pool. This will cause extra database load that needs to be assessed.

  • A naive solution using the OFFSET FETCH syntax to fetch sections of a table in individual queries will still cause table blocks to be scanned even though not all data is returned.

  • Is the table partitioned?

  • Are zone maps being used?

  • Maybe the real performance bottleneck cannot be solved by parallelism. Perhaps you have function based indexes that are being invoked for every row.

  • Is the data in the database spread across multiple spindles or is the one disk having to seek?

  • Is Exadata with storage indexes being used?

  • What is the impact of Python’s Global Interpreter Lock (GIL)? Maybe multiple Python processes should be used instead of threads.

  • What is the application doing with the data? Can the receiving end efficiently process it?

  • Is it better to execute a single query in Python but use a PARALLEL query hint? Or will this overload the database.

11.2. Database Round-trips

A round-trip is defined as the travel of a message from python-oracledb to the database and back. Calling each python-oracledb function, or accessing each attribute, will require zero or more round-trips. For example, inserting a simple row involves sending data to the database and getting a success response back. This is a round-trip. Along with tuning an application’s architecture and tuning its SQL statements, a general performance and scalability goal is to minimize round-trips because they impact application performance and overall system scalability.

Some general tips for reducing round-trips are:

11.2.1. Finding the Number of Round-Trips

Oracle’s Automatic Workload Repository (AWR) reports show ‘SQL*Net roundtrips to/from client’ and are useful for finding the overall behavior of a system.

Sometimes you may wish to find the number of round-trips used for a specific application. Snapshots of the V$SESSTAT view taken before and after doing some work can be used for this:

# Get the connection's session id
def get_session_id(connection):
    sql = "select sys_context('userenv','sid') from dual"
    result, = connection.cursor().execute(sql).fetchone()
    return result

 # Get the number of round-trips a session has made so far
 def get_round_trips(systemconn, sid):
     sql = """select
                  ss.value
              from
                  v$sesstat  ss,
                  v$statname sn
              where
                  ss.sid = :sid
                  and ss.statistic# = sn.statistic#
                  and sn.name like '%roundtrip%client%'"""
     round_trips, = systemconn.cursor().execute(sql, [sid]).fetchone()
     return round_trips


systemconn = oracledb.connect(user="system", password=spw, dsn=cs)
connection = oracledb.connect(user=un, password=pw, dsn=cs)

sid = get_session_id(connection)
round_trips_before = get_round_trips(systemconn, sid)

# Do some "work"
cursor.execute("select ...")
rows = cursor.fetchall()

round_trips_after = get_round_trips(systemconn, sid)

print(f"Round-trips required for query: {round_trips_after - round_trips_before}")

Note that V$SESSTAT is not accurate for pipelined database operations.

11.3. Statement Caching

Python-oracledb’s Cursor.execute() and Cursor.executemany() methods use statement caching to make re-execution of statements efficient. Statement caching lets Oracle Database cursors be used without re-parsing the statement. Statement caching also reduces metadata transfer costs between python-oracledb and the database. Performance and scalability are improved.

The python-oracledb Thick mode uses Oracle Call Interface statement caching, whereas the Thin mode uses its own implementation.

Each standalone or pooled connection has its own cache of statements with a default size of 20. The default size of the statement cache can be changed using the oracledb.defaults.stmtcachesize attribute. The size can be set when creating connection pools or standalone connections. In general, set the statement cache size to the size of the working set of statements being executed by the application. To manually tune the cache, monitor the general application load and the Automatic Workload Repository (AWR) “bytes sent via SQL*Net to client” values. The latter statistic should benefit from not shipping statement metadata to python-oracledb. Adjust the statement cache size to your satisfaction. With Oracle Database 12c (or later), the Thick mode statement cache size can be automatically tuned using an oraaccess.xml file.

11.3.1. Setting the Statement Cache

The statement cache size can be set globally with oracledb.defaults.stmtcachesize:

import oracledb

oracledb.defaults.stmtcachesize = 40

The value can be overridden in an oracledb.connect() call, or when creating a pool with oracledb.create_pool(). For example:

oracledb.create_pool(user="scott", password=userpwd, dsn="dbhost.example.com/orclpb",
                     min=2, max=5, increment=1, stmtcachesize=50)

When python-oracledb Thick mode uses Oracle Client 21 (or later), changing the cache size with ConnectionPool.reconfigure() does not immediately affect connections previously acquired and currently in use. When those connections are subsequently released to the pool and re-acquired, they will then use the new value. When the Thick mode uses Oracle Client prior to version 21, changing the pool’s statement cache size has no effect on connections that already exist in the pool but will affect new connections that are subsequently created, for example when the pool grows.

11.3.2. Tuning the Statement Cache

In general, set the statement cache to the size of the working set of statements being executed by the application. SODA internally makes SQL calls, so tuning the cache is also beneficial for SODA applications.

In python-oracledb Thick mode, the statement cache size can be automatically tuned with the Oracle Client Configuration oraaccess.xml file.

For manual tuning use views like V$SYSSTAT:

SELECT value FROM V$SYSSTAT WHERE name = 'parse count (total)'

Find the value before and after running application load to give the number of statement parses during the load test. Alter the statement cache size and repeat the test until you find a minimal number of parses.

If you have Automatic Workload Repository (AWR) reports you can monitor general application load and the “bytes sent via SQL*Net to client” values. The latter statistic should benefit from not shipping statement metadata to python-oracledb. Adjust the statement cache size and re-run the test to find the best cache size.

11.3.3. Disabling the Statement Cache

Statement caching can be disabled by setting the cache size to 0:

oracledb.defaults.stmtcachesize = 0

Disabling the cache may be beneficial when the quantity or order of statements causes cache entries to be flushed before they get a chance to be reused. For example if there are more distinct statements than cache slots, and the order of statement execution causes older statements to be flushed from the cache before the statements are re-executed.

Disabling the statement cache may also be helpful in test and development environments. The statement cache can become invalid if connections remain open and database schema objects are recreated. Applications can then receive errors such as ORA-3106. However, after a statement execution error is returned once to the application, python-oracledb automatically drops that statement from the cache. This lets subsequent re-executions of the statement on that connection to succeed.

When it is inconvenient to pass statement text through an application, the Cursor.prepare() call can be used to avoid statement re-parsing. If the cache_statement parameter in the Cursor.prepare() method is True and the statement cache size is greater than 0, then the statements will be added to the cache, if not already present. If the cache_statement parameter in the Cursor.prepare() method is False and the statement cache size is greater than 0, then the statement will be removed from the statement cache (if present) or will not be cached (if not present). The subsequent execute() calls use the value None instead of the SQL text.

This feature can prevent a rarely executed statement from flushing a potential more frequently executed one from a full cache. For example, if a statement will only ever be executed once:

cursor.prepare("select user from dual", cache_statement = False)
cursor.execute(None)

Alternatively,

sql = "select user from dual"
cursor.prepare(sql, cache_statement=False)
cursor.execute(sql)

Statements passed to prepare() are also stored in the statement cache.

11.4. Client Result Caching (CRC)

Python-oracledb applications can use Oracle Database’s Client Result Cache. The CRC enables client-side caching of SELECT query results in client memory for immediate use when the same query is re-executed. This is useful for reducing the cost of queries for small, mostly static, lookup tables, such as for postal codes. CRC reduces network round-trips, and also reduces database server CPU usage.

Note

Client Result Caching is only supported in python-oracledb Thick mode. See Enabling python-oracledb Thick mode.

The cache is at the application process level. Access and invalidation is managed by the Oracle Client libraries. This removes the need for extra application logic, or external utilities, to implement a cache.

CRC can be enabled by setting the database parameters CLIENT_RESULT_CACHE_SIZE and CLIENT_RESULT_CACHE_LAG, and then restarting the database, for example:

SQL> ALTER SYSTEM SET CLIENT_RESULT_CACHE_LAG = 3000 SCOPE=SPFILE;
SQL> ALTER SYSTEM SET CLIENT_RESULT_CACHE_SIZE = 64K SCOPE=SPFILE;
SQL> STARTUP FORCE

CRC can alternatively be configured in an oraaccess.xml or sqlnet.ora file on the Python host, see Client Configuration Parameters.

Tables can then be created, or altered, so repeated queries use CRC. This allows existing applications to use CRC without needing modification. For example:

SQL> CREATE TABLE cities (id number, name varchar2(40)) RESULT_CACHE (MODE FORCE);
SQL> ALTER TABLE locations RESULT_CACHE (MODE FORCE);

Alternatively, hints can be used in SQL statements. For example:

SELECT /*+ result_cache */ postal_code FROM locations