1. Introduction

This blog talks about a high level description of the mechanism behind PostgreSQL to execute an INSERT query. This process involves many steps of processing before the data is put in the right place. These process normally involves several catalog cache lookup to determine if the destination table exists or several checking on the constraint violations..etc. This blog will mainly focus on the part where the processing handle is passed to the PostgreSQL’s table access method API and its interaction with buffer manager and WAL routines. This is also the core area where an INSERT query is actually executed. If you are a developer looking to understand how PostgreSQL works internally, this blog may be helpful to you…

2. Table Access Method APIs in PostgreSQL

Pluggable table access method API has been made available since PostgreSQL 12, which allows a developer to redefine how PostgreSQL stores / retrieves table data. This API contains a total of 42 routines that need to be implemented in order to complete the implementation and honestly it is no easy task to understand all of them and to implement them. This API structure is defined in tableam.h under the name typedef struct TableAmRoutine

Today I will describe the routines related to INSERT.

3. INSERT Query Overall Call Flow

A few of the 42 routines will be called by executor just to complete an INSERT query. This section will describe these routines in the order they are called.

3.1 slot_callbacks

const TupleTableSlotOps *(*slot_callbacks) (Relation rel);

The executor needs to find out which set of tuple table slot (TTS) callback operation this table access method is compatible with. TTS is a set of routines that ensures the tuple storage is compatible between the executor and your access method. The executor will execute the TTS callback to translate your tuple strucuture to TupleTableSlot format in which the executor will understand. The default heap access method uses TTSOpsBufferHeapTuple defined in execTuples.c to handle this operation

3.2 heap_insert

void
heap_insert(Relation relation, HeapTuple tup, CommandId cid,
			int options, BulkInsertState bistate)

heap_insert is the entry point to perform the actual data insertion and it will undergo several other routines provided by buffer manager and WAL module in order to complete the insertion.

heap_prepare_insert

static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
									 TransactionId xid, CommandId cid, int options);

This is a subroutine for heap_insert where it will initialize the tuple header contents such as relation OID, infomasks, xmin, xmax values. It will also determine if the tuple is too big that TOAST is required to complete the insertion. These terms and parameters are very technical in PostgreSQL. If you are not sure what exactly they are, you could refer to resources here and here.

RelationGetBufferForTuple

extern Buffer RelationGetBufferForTuple(Relation relation, Size len,
										Buffer otherBuffer, int options,
										BulkInsertStateData *bistate,
										Buffer *vmbuffer, Buffer *vmbuffer_other);

This is an entry function to access buffer manager resources and all it is doing is ask the buffer manager to return a buffer ID that can be used to store the target tuple. This may sound very straightforward, but there is quite a lot of processing on the buffer manager side to properly determine a desired buffer location.

First, it will do a quick size check. If the input tuple is larger than the size of each buffer block, it will return immediately with error as TOAST has to be used in this case. Then it will try to put the tuple on the same page the system last inserted the tuple on to see if it will fit there. If not, it will utilize the free space map to find another page that could fit tuple. If that does not work out, then buffer manage will allocate a new data page (also referred to as extend) to be used to hold this new tuple. As soon as we have a desired buffer page determined, buffer manager will cache this page in the relation structure such that next time the same relation visits the buffer manager, it knows immediately about the reference to the last inserted block.

RelationPutHeapTuple

extern void RelationPutHeapTuple(Relation relation, Buffer buffer,
								 HeapTuple tuple, bool token);

Once we have identified the location of the buffer to store the tuple, the insert routine will then call RelationPutHeapTuple to actually put the tuple in the specified buffer location. This routine will again ask the buffer manager to get a pointer reference to the data page using the buffer ID we obtained from RelationGetBufferForTuple, then add the tuple data using PageAddItem() routine. Internally in buffer manager, it manages the relationship between buffer ID, buffer descriptor and the actual pointer to the data page to help us correctly identify and write to a data page. After a successful write, the routine will save a CTID of the inserted tuple. This ID is the location of this tuple and it consists of the data page number and the offset. For more information about how buffer manager works, you can refer to the resource here

Mark buffer dirty

extern void MarkBufferDirty(Buffer buffer);

At this point, the tuple data is already stored in the buffer manager referenced by a particular data page plus an offset, but it is not yet flushed to disk yet. In this case, we almost always will have to call MarkBufferDirty function to signal buffer manager that there are some tuples on the page that have not been flushed to disk and therefore in the next checkpoint, it will ensure the new tuples are flushed to disk.

[Insert WAL Record]

Last but not least, after doing all the hassle of finding a buffer location to put our tuple in and mark it as dirty, it is time for the heap_insert routine to populate a WAL record. This part is not the focus of this blog so I will skip the high level details of WAL writing.

3.3 End of the insertion

At this point the insertion of a new tuple data has finished and proper WAL record has been written. The routine will once again save the CTID value that we derived during the data insertion and save this value to the TTS structure so the executor also gets a copy of the location of the tuple. Then it will clean up the local resources before returning.

4. Summary

What we have discussed here is the basic call flow of a simple sequential scan. If we were to visualize the process, it should look something like this:

1. Introduction

There are many approaches for PostgreSQL to retrieve the data back to the user. Depending on the user’s input query, the planner module is responsible for selecting the most optimum approach to retrieve the requested data. Sequential scan is one of these approaches that is mostly selected when the user requests a large volume of data (for example, “SELECT * from tablename;”) or when a table has no index declared. Sequential scan is mostly handled by the table access method API within PostgreSQL and heap access method is the default one PostgreSQL uses today. In this short post, I will show you how sequential scan is done in the table access method API.

2. Table Access Method APIs in PostgreSQL

Pluggable table access method API has been made available since PostgreSQL 12, which allows a developer to redefine how PostgreSQL stores / retrieves table data. This API contains a total of 42 routines that need to be implemented in order to complete the implementation and honestly it is no easy task to understand all of them and to implement them. This API structure is defined in tableam.h under the name typedef struct TableAmRoutine

Today I will describe the routines related to sequential scan and I hope it could help you if you are someone looking to create your own table access method.

3. Sequential Scan Overall Call Flow

Few of the 42 routines will be called by executor just to complete a sequential scan request. This section will describe these routines in the order they are called.

3.1 relation_size

uint64		(*relation_size) (Relation rel, ForkNumber forkNumber);

relation_size is the first routine to be called and it is relatively simple. The caller will expect the routine to return the total size of the relation described by rel and forkNumber. The default heap access method will simply invoke the storage manager smgr to find the number of data blocks this particular relation physically occupies on disk and multiplies that number with the size of each block BLCKSZ (default is 8k). If you are not sure about the relationship between relation and its fork number, you could refer to this blog to get more information.

The size returned by this routine basically sets the boundary of our sequential scan.

3.2 slot_callbacks

const TupleTableSlotOps *(*slot_callbacks) (Relation rel);

Next, the executor needs to find out which set of tuple table slot (TTS) callback operation this table access method is compatible with. TTS is a set of routines that ensures the tuple storage is compatible between the executor and your access method. The executor will execute the TTS callback to translate your tuple strucuture to TupleTableSlot format in which the executor will understand. The default heap access method uses TTSOpsBufferHeapTuple defined in execTuples.c to handle this operation

3.3 scan_begin

TableScanDesc (*scan_begin) (Relation rel,
							 Snapshot snapshot,
							 int nkeys, struct ScanKeyData *key,
							 ParallelTableScanDesc pscan,
							 uint32 flags);

Now the scan can officially begin. This is sequential scan’s initialization routine in which it will allocate a new scan descriptor using the parameters passed in by the executor. The purpose of scan descriptor structure is to keep track of the sequential scan while it is being executed. For example, to track where the scan should begin,; when was the block number of the last scan; which block should we resume scanning and how many blocks have been scanned…etc. scan descriptor will be destroyed once the sequential scan has completed.

The executor expects the routine to return a fully allocated and initialize pointer to TableScanDesc struct

3.4 scan_getnextslot

bool		(*scan_getnextslot) (TableScanDesc scan,
								 ScanDirection direction,
								 TupleTableSlot *slot);

This is the main routine for sequential scan where the caller expects the routine to fetch one tuple from the buffer manager, converts it to the TTS format in which executor understands and save it in the input pointer called slot. Each call to this routine will results in one tuple to be returned. If a table contains 1000 tuples, this function will be called 1000 times. The boolean return code is the indication to the caller if the routine has more tuples to return, as soon as false is returned, it signals the executor that we have exhausted all the tuples and it should stop calling this function.

In normal sequential scan case, this routine works in per-page mode. This means it will read one full block from buffer manager and scan it to get all the tuple addresses and their offsets in the scan descriptor, so in the subsequent calls to the same function, it will not load the full page again from buffer manage all the time; it will only start to load the next block when all the tuples on the current block have been scanned and returned.

As you can see, the scan descriptor plays an important role here as most of the control information is saved there and is regularly updated whenever a call the scan_getnextslot is made.

3.5 scan_end

void		(*scan_end) (TableScanDesc scan);

This is the last routine to be called to basically clean up the table scan descriptor, which was used heavily during the sequential scan. At this point, the executor should already have all the tuple information from the sequential scan methods.

4. Prepare the Data to Return

Now, the executor is finished with the table access method and has already had access to all the tuples for a particular relation. It then needs to go through another round of filtering to determine which of these tuples satisfy the condition set by the user, (for example, when user gives the WHERE clause to limit the scan results). This is done in another infinite for loop in execScan.c to perform ExecQual on each TTS. Finally, the end results will be sent to the end user.

5. Summary

What we have discussed here is the basic call flow of a simple sequential scan. If we were to visualize the process, it should look something like this:

1. Introduction

This blog is to follow up on the post I published back in July, 2020 about achieving an in-memory table storage using PostgreSQL’s pluggable storage API. In the past few months, my team and I have made some progress and did a few POC patches to prove some of the unknowns and hypothesis and today I would like to share our progress.

2. The PostgreSQL Buffer Manager

In the previous post, I mentioned that we would like to build a new in-memory based storage that is based on the existing buffer manager and its related component and hooked it up with the pluggable storage API. To achieve this, my team and I underwent an in-depth study to understand how the current buffer manager works in PostgreSQL and this chapter at interdb.jp is a good starting point for us to gain a general understanding of the buffer manager design in good details.

The current PostgreSQL buffer manager follows a 3-layer buffer design to manage the data pages as illustrated by this image below:

where it consists of

• Buffer Table (hash table)
• Buffer Descriptors (Array)
• Buffer Pool (Array)

2.1 Page Table

Buffer Table is used like a routing table between PostgreSQL Core and the buffer manager. It is managed using the existing hash table utilities and uses buffer_tag to look up the page descriptor and buffer id. Buffer_tag is a structure that contains the table space, database, table name.

2.2 Buffer Descriptor

Buffer Descriptor is used to store the status of a buffer block and also the content lock. Refcount is a part of the buffer state, will be used to indicate the insert and delete operation. it will be increased by one when there is an insertion, and decreased by one when there is a deletion. The Vacuum process will reclaim this page once refcount reaches to 0.

2.3 Buffer Pool

Buffer Pool has a one to one relationship with buffer descriptor. it can be treated a simple pointer pointing to the beginning of the buffer pool, each buffer pool slot is defined as 8KB for now. This is the lowest layer in the buffer manager structure before a page is flushed to disk. The BM_DIRTY status flag is used to indicate if a page in the buffer pool is to be flushed to disk

In addition to buffer pool, buffer manager also utilizes a ring buffer for reading and writing a huge table whose size exceeds 1/4 of the buffer pool size. Clock Sweep algorithm is used to find a victim page in the ring buffer to eject and flush to disk so new page can enter, thus the name, ring buffer.

3. The In-Memory Only Buffer Manager

Having a general understanding of the existing buffer manager’s strucutre, we hypothesize that we could potentially improve its IO performance by eliminating the need to flush any buffer data to disk. This means that the in-memory only version of buffer manager itself is the storage media. For this reason, its strucutre can be simplified as:

where the buffer descriptor points to a dedicated memory storage that contains the actual page and tuple. This memory space can be allocated to a certain size at initlaization and there will not be a need to flush a page to disk. All data page and tuple will reside in this memory space. In the case where there is a huge reading and writing load, the ring buffer will not be allocated as the logic to find a victim page to evict and flush to disk will be removed since everything will reside in a dedicated memory space. For this reason, if the memory space is not sufficiently allocated, the user will get “no unpin buffer” is available, which basically means “your disk is full” and you need to do delete and vacuum.

Using this approach, when the server shuts down or restarts, the data in this memory space is of course lost. So, data persistence to disk would be a topic of interest next, but right now, we already see some useful business case with this in-memory based table where data processing speed is more important than data persistence

4. Initial Results

Using the same tuple structure and logic as the current heap plus the memory based buffer manager with 1GB of memory allocated, we observe some interesting increase in performance comparing to PostgreSQL with default settings. For 20 million row, we observe about 50% increase for insert, 70% increase for update, 60% increase for delete and 30% increase in vacuum. This result is not too bad considering we are still at the early stages and I am sure there are many other ways to make it even faster

5. Next Step

Having some solid results from the initial test, it would make sense for us to also be looking into having the index tuples as in-memory storage only. In addition, free space map and visibility map files that are stored in the PG cluster directory could potentially also be made as in-memory to possibly further increase the DML performance.