Skip to content

Iterator

Jay Huh edited this page Jun 17, 2024 · 32 revisions

Introduction

All data in the database is logically arranged in sorted order. An application can specify a key comparison method that specifies a total ordering of keys. An Iterator API allows an application to do a range scan on the database. The Iterator can seek to a specified key and then the application can start scanning one key at a time from that point. The Iterator API can also be used to do a reverse iteration of the keys in the database. A consistent-point-in-time view of the database is created when the Iterator is created. Thus, all keys returned via the Iterator are from a consistent view of the database.

Consistent View

If ReadOptions.snapshot is given, the iterator will return data as of the snapshot. If it is nullptr, the iterator will read from an implicit snapshot as of the time the iterator is created. The implicit snapshot is preserved by pinning resource. There is no way to convert this implicit snapshot to an explicit snapshot.

Error Handling

Iterator::status() returns the error of the iterating. The errors include I/O errors, checksum mismatch, unsupported operations, internal errors, or other errors.

If there is no error, the status is Status::OK(). If the status is not OK, the iterator will be invalidated too. In another word, if Iterator::Valid() is true, status() is guaranteed to be OK() so it's safe to proceed other operations without checking status():

for (it->Seek("hello"); it->Valid(); it->Next()) {
  // Do something with it->key() and it->value().
}
if (!it->status().ok()) {
  // Handle error. it->status().ToString() contains error message.
}

On the other hand, if Iterator::Valid() is false, there are two possibilities: (1) We reached the end of the data. In this case, status() is OK(); (2) there is an error. In this case status() is not OK(). It is always a good practice to check status() if the iterator is invalidated.

Seek() and SeekForPrev() discard previous status.

Note that in release 5.13.x or earlier (before https://github.com/facebook/rocksdb/pull/3810 which was merged on May 17, 2018) the behavior of status() and Valid() used to be different:

  • Valid() could return true even if status() is not ok. This could sometimes be used to skip over corrupted data. This is not supported anymore. The intended way of dealing with corrupted data is RepairDB() (see db.h).
  • Seek() and SeekForPrev() didn't always discard previous status. Next() and Prev() didn't always preserve non-ok status.

Iterating upper bound and lower bound

A user can specify an upper bound of your range query by setting ReadOptions.iterate_upper_bound for the read option passed to NewIterator(). By setting this option, RocksDB doesn't have to find the next key after the upper bound. In some cases, some I/Os or computation can be avoided. In some specific workloads, the improvement can be significant. Note it applies to both of forward and backward iterating. The behavior is not defined when you do SeekForPrev() with a seek key higher than upper bound, or calling SeekToLast() with the last key to be higher than an iterator upper bound, although RocksDB will not crash.

Similarly, ReadOptions.iterate_lower_bound can be used with backward iterating to help RocksDB optimize the performance.

See the comment of the options for more information.

Resource pinned by iterators and iterator refreshing

Iterators by themselves don't use much memory, but it can prevent some resource from being released. This includes:

  1. memtables and SST files as of the creation time of the iterators. Even if some memtables and SST files are removed after flush or compaction, they are still preserved if an iterator pinned them.
  2. data blocks for the current iterating position. These blocks will be kept in memory, either pinned in block cache, or in the heap if block cache is not set. Please note that although normally blocks are small, in some extreme cases, a single block can be quite large, if the value size is very large.

So the best use of iterator is to keep it short-lived, so that these resource is freed timely.

An iterator has some creation costs. In some use cases (especially memory-only cases), people want to avoid the creation costs of iterators by reusing iterators. When you are doing it, be aware that in case an iterator getting stale, it can block resource from being released. So make sure you destroy or refresh them if they are not used after some time, e.g. one second. When you need to treat this stale iterator, before release 5.7, you'll need to destroy the iterator and recreate it if needed. Since release 5.7, you can call an API Iterator::Refresh() to refresh it. By calling this function, the iterator is refreshed to represent the recent states, and the stale resource pinned previously is released.

Prefix Iterating

Prefix iterator allows users to use bloom filter or hash index in iterator, in order to improve the performance. However, the feature has limitation and may return wrong results without reporting an error if misused. So we recommend you to use this feature carefully. For how to use the feature, see Prefix Seek. Options total_order_seek and prefix_same_as_start are only applicable in prefix iterating.

Read-ahead

RocksDB does automatic readahead and prefetches data on noticing more than 2 IOs for the same table file during iteration. This applies only to the block based table format. The readahead size starts with 8KB and is exponentially increased on each additional sequential IO, up to a max of BlockBasedTableOptions.max_auto_readahead_size (default 256 KB). This helps in cutting down the number of IOs needed to complete the range scan. This automatic readahead is enabled only when ReadOptions.readahead_size = 0 (default value). On Linux, readahead syscall is used in Buffered IO mode, and an AlignedBuffer is used in Direct IO mode to store the prefetched data. (Automatic iterator-readahead is available starting 5.12 for buffered IO and 5.15 for direct IO).

If your entire use case is dominated by iterating and you are relying on OS page cache (i.e using buffered IO), you can choose to turn on page cache's internal readahead by having DBOptions.advise_random_on_open = false. This is more helpful if you run on hard drives or remote storage, but may not have much actual effects on directly attached SSD devices.

ReadOptions.readahead_size provides read-ahead support in RocksDB for very limited use cases. The limitation of this feature is that, if turned on, the constant cost of the iterator will be much higher. So you should only use it if you iterate a very large range of data, and can't work it around using other approaches. A typical use case will be that the storage is remote storage with very long latency, OS page cache is not available and a large amount of data will be scanned. By enabling this feature, every read of SST files will read-ahead data according to this setting. Note that one iterator can open each file per level, as well as all L0 files at the same time. You need to budget your read-ahead memory for them. And the memory used by the read-ahead buffer can't be tracked automatically.

The readahead, both automatic and fixed with ReadOptions.readahead_size, can be done asynchronously by RocksDB if the ReadOptions.async_io flag is set. In this case, the iterator will try to prefetch data without blocking the iterating thread. See Asynchronous IO for more details.

We are looking for improving read-ahead in RocksDB.

Multi Column Family Iterator (Experimental)

MultiCfIterator (CoalescingIterator and AttributeGroupIterator) allows users to iterate over keys across multiple column families. It supports most of the existing iterator features. For how to use the feature, see Multi Column Family Iterator.

Contents

Clone this wiki locally